JIST - Society for Imaging Science and Technology

JIST 

Vol. 51, No. 1 

January/February 

2007 

Journal of 

Imaging Science 

and Technology 

imaging.org 

Society for Imaging Science and Technology

Editorial Staff 

Melville Sahyun, editor 

sahyun@infionline.net 

Donna Smith, production manager 

dsmith@imaging.org 

Editorial Board 

Philip Laplante, associate editor 

Michael Lee, associate editor 

Nathan Moroney, associate editor 

Mitchell Rosen, color science editor 

David S. Weiss, associate editor 

David R. Whitcomb, associate editor 

JIST papers are available for purchase 

at www.imaging.org and through 

ProQuest. They are indexed in 

INSPEC, Chemical Abstracts, Imaging 

Abstracts, COMPENDEX, and ISI: 

Science Citation Index. 

Orders for subscriptions or single 

copies, claims for missing numbers, 

and notices of change of address 

should be sent to IS&T via one of the 

means listed below. 

IS&T is not responsible for the accuracy 

of statements made by authors and 

does not necessarily subscribe to their 

views. 

Copyright ©2007, Society for Imaging 

Science and Technology. Copying 

of materials in this journal for internal 

or personal use, or the internal or personal 

use of specific clients, beyond 

the fair use provisions granted by the 

US Copyright Law is authorized by 

IS&T subject to payment of copying 

fees. The Transactional Reporting Service 

base fee for this journal should be 

paid directly to the Copyright Clearance 

Center (CCC), Customer Service, 

508/750-8400, 222 Rosewood Dr., 

Danvers, MA 01923 or online at 

www.copyright.com. Other copying 

for republication, resale, advertising or 

promotion, or any form of systematic 

or multiple reproduction of any material 

in this journal is prohibited except with 

permission of the publisher. 

Library of Congress Catalog Card 

No. 59-52172 

Printed in the USA. 

Society for Imaging Science and 

Technology 

7003 Kilworth Lane 

Springfield, VA 22151 

www.imaging.org 

info@imaging.org 

703/642-9090 

703/642-9094 fax 

Manuscripts should be sent to the 

postal address above as describe at 

right. E-mail PDF and other files as requested 

to dsmith@imaging.org. 

Guide for Authors 

Scope: The Journal of Imaging Science and Technology (JIST) is dedicated to the advancement of imaging science knowledge, the 

practical applications of such knowledge, and how imaging science relates to other fields of study. The pages of this journal are 

open to reports of new theoretical or experimental results, and to comprehensive reviews. Only original manuscripts that have not 

been previously published nor currently submitted for publication elsewhere should be submitted. Prior publication does not refer 

to conference abstracts, paper summaries, or non-reviewed proceedings, but it is expected that Journal articles will expand in scope 

the presentation of such preliminary communication. Please include keywords on your title and abstract page. 

Editorial Process/Submission of Papers for Review: All submitted manuscripts are subject to peer review. (If a manuscript appears better 

suited to publication in the Journal of Electronic Imaging, published jointly by IS&T and SPIE, the editor will make this recommendation.) 

To expedite the peer review process, please recommend two or three competent, independent reviewers. The editorial staff, will 

take these under consideration, but is not obligated to use them. 

Manuscript Guidelines: Please follow these guidelines when preparing accepted manuscripts for submission. 

• Manuscripts should be double-spaced, single-column, and numbered. It is the responsibility of the author to prepare a succinct, 

well-written, paper composed in proper English. JIST generally follows the guidelines found in the AIP Style Manual, available 

from the American Institute of Physics. 

• Documents may be created in Microsoft Word, WordPerfect, or LaTeX/REVTeX. 

• Manuscripts must contain a title page that lists the paper title, full name(s) of the author(s), and complete affiliation/address for 

each author. Include an abstract that summarizes objectives, methodology, results, and their significance; 150 words maximum. 

Provide at least four key words. 

• Figures should conform to the standards set forth at www.aip.org/epub/submitgraph.html. 

• Equations should be numbered sequentially with Arabic numerals in parentheses at the right margin. Be sure to define symbols 

that might be confused (such as ell/one, nu/vee, omega/w). 

• For symbols, units, and abbreviations, use SI units (and their standard abbreviations) and metric numbers. Symbols, acronyms, 

etc., should be defined on their first occurrence. 

• Illustrations: Number all figures, graphs, etc. consecutively and provide captions. Figures should be created in such a way that 

they remain legible when reduced, usually to single column width (3.3 inches/8.4 cm); see also 

www.aip.org/epub/submitgraph.html for guidance. Illustrations must be submitted as .tif or .eps files at full size and 600 dpi; 

grayscale and color images should be at 300 dpi. JIST does not accept .gif or .jpeg files. Original hardcopy graphics may be sent 

for processing by AIP, the production house for JIST. (See note below on color and supplemental illustrations.) 

• References should be numbered sequentially as citations appear in the text, format as superscripts, and list at the end of the document 

using the following formats: 

• Journal articles: Author(s) [first/middle name/initial(s), last name], “title of article (optional),” journal name (in italics), ISSN 

number (e.g. for JIST citation, ISSN: 1062-3701), volume (bold): first page number, year (in parentheses). 

• Books: Author(s) [first/ middle name/initial(s), last name], title (in italics), (publisher, city, and year in parentheses) page reference. 

Conference proceedings are normally cited in the Book format, including publisher and city of publication (Springfield, VA, for all 

IS&T conferences), which is often different from the conference venue. 

• Examples 

1. H. P. Le, Progress and trends in ink-jet printing technology, J. Imaging Sci. Technol. 42, 46 (1998). 

2. E. M. Williams, The Physics and Technology of Xerographic Processes (John Wiley and Sons, New York, 1984) p. 30. 

3. Gary K. Starkweather, “Printing technologies for images, gray scale, and color,” Proc. SPIE 1458: 120 (1991). 

4. Linda T. Creagh, “Applications in commercial printing for hot melt ink-jets,” Proc. IS&T’s 10th Int’l. Congress on Adv. In 

Non-Impact Printing Technologies (IS&T, Springfield, VA 1994) pp. 446-448. 

5. ISO 13655-1996 Graphic technology: Spectral measurement and colorimetric computation for graphic arts images (ISO, 

Geneva), www.iso.org. 

6. Society for Imaging Science and Technology website, www.imaging.org, accessed October 2003. 

Reproduction of Color: Authors who wish to have color figures published in the print journal will incur color printing charges. 

The cost for reproducing color illustrations is $490 per page; color is not available to those given page waivers, nor can color page 

charges be negotiated or waived. Authors may also choose to have their figures appear in color online and in grayscale in the printed 

journal. There is no additional charge for this, however those who choose this option are responsible for ensuring that the captions 

and descriptions in the text are readable in both color and black-and-white as the same file will be used in the online and 

print versions of the journal. Only figures saved as TIFF/TIF or EPS files will be accepted for posting. Color illustrations may be 

also submitted as supplemental material for posting on the IS&T website for a flat fee of $100 for up to five files. 

Website Posting of Supplemental Materials: Authors may also submit additional (supplemental) materials related to their articles 

for posting on the IS&T Website. Examples of such materials are charts, graphs, illustrations, or movies that further explain the 

science or technology discussed in the paper. Supplemental materials will be posted for a flat fee of $100 for up to five files. For 

each additional file, a $25 fee will be charged. Fees must be received before supplemental materials will be posted. As a matter of 

editorial policy, appendices are normally treated as supplemental material. 

Submission of Accepted Manuscripts: Author(s) will receive notification of acceptance (or rejection) and reviewers’ 

reports. Those whose manuscripts have been accepted for publication will receive correspondence informing them of the issue for 

which the paper is tentatively scheduled, links to copyright and page charge forms, and detailed instructions for submitting accepted 

manuscripts. A duly signed transfer of copyright agreement form is required for 

publication in this journal. No claim is made to original US Government works. 

Page charges: Page charges for the Journal is $80/printed page. Such payment is 

not a condition for publication, and in some circumstances page charges are 

waived. Requests for waivers must be made in writing to the managing editor prior 

to acceptance of the paper and at the time of submission. 

Manuscripts submissions: Manuscripts should be submitted both electronically 

and as hardcopy. To submit electronically, send a single PDF file attached to an e- 

mail message/cover letter to jist@imaging.org. To submit hardcopy, mail 2 singlespaced, 

single-sided copies of the manuscript to: IS&T. With both types of submission, 

include a cover letter that states the paper title; lists all authors, with complete 

contact information for each (affiliation, full address, phone, fax, and e-mail); identifies 

the corresponding author; and notes any special requests. Unless otherwise 

stated, submission of a manuscript will be understood to mean that the paper has 

been neither copyrighted, classified, or published, nor is being considered for 

publication elsewhere. Authors of papers published in the Journal of Imaging 

Science and Technology are jointly responsible for their content. Credit for the 

content and responsibility for errors or fraud are borne equally by all authors. 

JOURNAL OF IMAGING SCIENCE AND TECH- 

NOLOGY ( ISSN:1062-3701) is published bimonthly 

by The Society for Imaging Science and Technology, 

7003 Kilworth Lane, Springfield, VA 22151. Periodicals 

postage paid at Springfield, VA and at 

additional mailing offices. Printed in Virginia, 

USA. 

Society members may receive this journal as part of 

their membership. Forty-five dollars ($45.00) of 

membership dues are allocated to this subscription. 

IS&T members may refuse this subscription by written 

request. Domestic institution and individual nonmember 

subscriptions are $195/year or $50/single 

copy. The foreign subscription rate is $205/year. 

For online version information, contact IS&T. 

POSTMASTER: Send address changes to JOURNAL 

OF IMAGING SCIENCE AND TECHNOLOGY, 

7003 Kilworth Lane, Springfield, VA 22151.

JIST 

Vol. 51, No. 1 

January/February 

2007 

Journal of 

Imaging Science 

and Technology ® 

iii 

iv 

From the Editor 

To the Editor 

Feature Article 

1 Improved Pen Alignment for Bidirectional Printing 

Edgar Bernal, Jan P. Allebach, and Zygmunt Pizlo 

General Papers 

23 Characterization of Red-Green and Blue-Yellow Opponent Channels 

Bong-Sun Lee, Zygmunt Pizlo, and Jan P. Allebach 

34 High Dynamic Range Image Compression by Fast Integrated Surround 

Retinex Model 

Lijie Wang, Takahiko Horiuchi, and Hiroaki Kotera 

44 Illumination-Level Adaptive Color Reproduction Method with Lightness 

Adaptation and Flare Compensation for Mobile Display 

Myong-Young Lee, Chang-Hwan Son, Jong-Man Kim, 

Cheol-Hee Lee, and Yeong-Ho Ha 

53 Influence of Paper on Colorimetric Properties of an Ink Jet Print 

Marjeta Černi~ and Sabina Bra~ko 

61 Development of a Multi-spectral Scanner using LED Array for Digital 

Color Proof 

Shoji Yamamoto, Norimichi Tsumura, Toshiya Nakaguchi, and Yoichi 

Miyake 

70 Spectral Color Imaging System for Estimating Spectral Reflectance 

of Paint 

Vladimir Bochko, Norimichi Tsumura, and Yoichi Miyake 

79 Digital Watermarking of Spectral Images Using PCA-SVD 

Long Ma, Changjun Li, and Shuni Song 

86 Qualification of a Layered Security Print Deterrent 

Steven J. Simske and Jason S. Aronoff 

continued on next page 

imaging.org 

Society for Imaging Science and Technology

IS&T BOARD OF DIRECTORS 

President 

James R. Milch Jim 

Director Medical Sciences 

Eastman Kodak Company 

continued from previous page 

96 Preparation of Gold Nanoparticles in a Gelatin Layer Film Using Photographic 

Materials (5): Characteristics of Gold Nanoparticles Prepared on an Ultrafine Grain 

Photographic Emulsion 

Ken’ichi Kuge, Tomoaki Nakao, Seiji Saito, Ohiro Hikosaka, and Akira Hasegawa 

Immediate Past President 

James C. King Jim 

Principal Scientist 

Adobe Systems Incorporated 

Executive Vice President 

Eric G. Hanson 

Department Manager 

Hewlett Packard Company 

Conference Vice President 

Rita Hofmann 

Chemist, R&D Manager 

Ilford Imaging Switzerland GmbH 

Publication Vice President 

Franziska Frey 

Assist. Prof., School of Print Media 

Rochester Institute of Technology 

Secretary 

Ramon Borrell 

Technology Strategy Director 

Hewlett Packard Company 

Treasurer 

Peter D. Burns 

Principal Scientist 


Vice Presidents 

Stefi Baum 

Director, Chester F. Carleson Center 

for Imaging Science 

Laura Kitzmann 

Marketing Dev. & Comm. Manager 

Sensient Imaging Technologies, Inc. 

Michael A. Kriss 

Retired 

Howard A. Mizes 

Principle Scientist, Xerox Corp. 

Jin Mizuguchi 

Professor, Yokohama National Univ. 

David Weiss 

Scientist Fellow, NexPress Solutions, 

Inc. 

IS&T Conference Calendar 

For details and a complete listing of conferences, visit www.imaging.org 

Electronic Imaging 

IS&T/SPIE 19th Annual Symposium 

January 28–February 1, 2007 

San Jose, California 

General chairs: Michael A. Kriss 

and Robert A. Sprague 

International Symposium on Technologies for 

Digital Fulfillment 

March 3–March 5, 2007 

Las Vegas, Nevada 

General chair: Stuart Gordon 

Archiving 2007 

May 21–May 24, 2007 

Arlington, Virginia 

General chair: Scott Stovall 

Ninth International Symposium 

on Multispectral Color Science 

and Application cosponsored by IS&T 

May 30–June 1, 2007 

Taipe, Taiwan 

General chairs: Tien-Rien Lee and Yoichi Miyake 

Digital Fabrication Processes Conference 

September 16–September 20, 2007 

Anchorage, Alaska 

General chair: Ross Mills 

NIP23: The 23rd International Congress on 

Digital Printing Technologies 

September 16–September 20, 2007 

Anchorage, Alaska 

General chair: Ramon Borrell 

IS&T/SID’s Fifteenth Color Imaging 

Conference cosponsored by SID 

November 5–November 9, 2007 

Albuquerque, New Mexico 

General chairs: Jan Morovic 

and Charles Poynton 

Chapter Director 

Franziska Frey – Rochester 

Takashi Kitamura – Japan 

Executive Director 

Suzanne E. Grinnan 

IS&T Executive Director 

ii 

/

From the Editor 

Starting this year, on-line will be the default method for 

IS&T members to receive the Journal of Imaging Science and 

Technology. Many readers already see the Journal on line or 

simply download pdf’s of articles of interest. Those without 

subscriptions are more likely, as well, to see electronic versions 

of the Journal’s articles rather than the original print 

hardcopy. 

At the same time the number of articles which address 

color science, either as the principal focus or a peripheral 

issue, is increasing steadily as illustrated by the General Papers 

in this issue, eight of which deal with aspects of color 

imaging. The number of submissions with respect to all aspects 

of imaging science with color illustrations is likewise 

increasing; authors expect to be able to illustrate their technical 

articles using color. However, fewer authors or their 

institutions are willing to subsidize the high cost of color 

printing in order to reach the limited audience of the highquality 

print edition, and are therefore opting for color online 

exclusively. When accurate color in an illustration is 

critical to understanding the content of an article, we need 

to establish a process that ensures that our readers of soft 

copy versions see the color graphics authentically, in the way 

the author(s) intend. 

Accordingly our subscribers and readers should be enabled 

to colorimetrically interpret the color values that sit in 

our articles correctly so that they may be viewed appropriately 

on their displays and printers. Our Editorial Staff has 

chosen sRGB as the default color space at this time because 

it is necessary in our production workflow to use a color 

space that is compatible with both web and print publication; 

this selection is subject to future revision, i.e., it may be 

permanent only as long as there continues to be a demand 

for us to publish a print edition of the Journal. But until that 

time, readers of the on-line Journal should calibrate their 

displays and printers accordingly. We anticipate that authors 

in the future may request different color spaces for their 

illustrations; in that case they will be able to choose a nondefault 

space, but will need to supply the Journal with an 

appropriate ICC profile. 

These topics are treated in more detail in the following 

Letter to the Editor, which provides guidelines for submission, 

display, and printing of color imagery in a webpublished 

Journal. These guidelines have been kindly providedbyDr.PatrickHerzog. 

—M. R. V. Sahyun 

iii

Letter to the Editor: 

Guidelines for the Handling of Color in IS&T Journal Papers 

Driven by the nature of our Society, color has been one 

of the major topics of scientific papers published in this 

Journal, and it is quite clear that showing research results 

frequently has required reproducing color images in true 

color. In color reproduction, however, it is difficult to 

achieve the required level of color faithfulness, and in the 

past authors have not been able to afford paying the extra 

costs for color printing. 

Journal papers are now published electronically, and 

color printers and monitors have become ubiquitous, so the 

scenario is changing, allowing for an extensive use of color. 

We would like to encourage even more use of color images, 

graphics, etc., wherever meaningful for the subject, or wherever 

appropriate to support the clarity of the paper. However, 

handling color is still not simple in general, and its 

extensive use may incur additional problems. We intend 

these guidelines to keep such problems to a minimum. Accordingly, 

we foresee different work paths, and hope to provide 

solutions for authors with different levels of color 

experience. 

Author guidelines 

General rules 

1. Keep in mind that in the print version of your paper 

you may choose to have the figures printed in grayscale, 

which may also be the case when readers who 

do not have access to color printers, purchase, download, 

and print the PDF of your article. Hence, you 

should make sure that all color images or graphics 

also reproduce well in grayscale. 

2. The default color space is sRGB (according to ISO/ 

IEC 61966-2.1). Make sure that all figures, graphics, 

etc. have been prepared for this color space or converted 

into sRGB after preparation. 

Why sRGB? 

sRGB is the system color space in Windows 2000 and XP, 

and will be the default color space with Windows Vista. 

Mac OS does not generally assume sRGB, but has built-in 

color management capabilities, so that it will correctly display 

images and PDFs if they have appropriate profiles embedded. 

Moreover, most printer manufacturers assume RGB 

data to be sRGB if no color management is used, so that one 

should get reasonable results also on office printers. 

How to create sRGB data 

Non-Color Expert Level 

Color conversions are carried out by means of a color management 

system using color profiles (ICC profiles). If you do 

not know how to create the proper color transforms, you are 

unlikely to require high color fidelity. To create a file properly, 

follow the simple guidelines given here. 

Simple Guidelines for Achieving sRGB-like behavior on PCs 

If you are using Windows 2000, XP, or Vista, then your system’s 

default color space is sRGB. This is good news. Moreover, 

most monitor manufacturers have adopted the sRGB 

standard in a way that every monitor by default approximates 

sRGB. This approximation does not guarantee color 

accuracy, but it does ensure that the computer/monitor system 

is not entirely off in terms of color. In other words: If 

you display a graphic or image on your system, and you like 

what you see, then it is safe for you to claim that the data 

was prepared for sRGB. 

Important Note: A precondition is that you have not 

modified the entire system, neither in terms of the operating 

system’s color settings, nor in terms of the monitor. If you 

have modified color temperature, color balance, RGB gains, 

etc., the display most probably no longer follows an approximate 

sRGB state. In this case try to reset the monitor to its 

default settings using the onscreen menu. 

Simple Guidelines for Achieving sRGB-like behavior 

on Macs 

If you are using a Macintosh, the default setup is different 

than sRGB. The main difference is that the Mac uses a 

gamma of 1.8 instead of the 2.2 of sRGB. If you do not 

know how to convert color images into another color space, 

we recommend changing the default setting. Just open the 

Monitor Control Panel from the System Preferences menu, 

click on the “Colors” tab, and press the “Calibrate” button. 

If your monitor is in good shape, you can use the nonexpert 

mode. Set the gamma correction to “2.2 TV gamma,” 

and the desired color temperature to “D65” or “uncorrected.” 

Some monitors, e.g., LCD, provide a greater luminance 

range at a slightly lower color temperature, e.g., D60, 

which may, in that case, be a better compromise. Note: Do 

not perform this manual correction if you are using a monitor 

calibration system (see below). 

Making use of Monitor Calibration Tools 

If you are using monitor calibration hard- and/or software, 

the sRGB state may also be void. In this case, you should use 

color conversion software (e.g., Photoshop) to convert the 

data into sRGB using your up-to-date monitor profile as a 

source profile and the sRGB profile as destination profile. If 

you do not know how to do this, carry out the following 

instead: Rerun your monitor calibration system and choose 

an sRGB setting if possible, or at least set the target gamma 

to 2.2 and the white point to D65 (or 6500 K). 

iv

Summary 

To summarize, there are two ways to achieve the required 

sRGB based color data. The first one is to put the computer 

and monitor in a state where it approximates sRGB. In this 

case, you can take any color data as perceived on the monitor 

as valid. The second way is to leave the computer/ 

monitor as it is, and to have a valid color profile available. 

Prepare all the color data so that you are happy with them. 

Afterwards, use the monitor profile as source profile and 

sRGB as destination profile and convert the data into sRGB. 

Color Expert Level 

We assume that you know how to convert color data into 

sRGB. If in doubt, follow the guidelines for non-color experts. 

We recommend that all color images have the sRGB 

profile embedded. Do not use color spaces different than 

sRGB even if embedded profiles in principle allow for this: 

Many web browsers or PDF viewers may not support embedded 

profiles, hence display and print quality may be 

compromised. 

Color Plate Appendix 

If you feel that sRGB limits your color data unacceptably, 

there is the option to include an appendix of color plates, 

which will not appear in the published paper version of the 

journal, but will be available as Supplemental Material on 

the IS&T website. For these color plates you can use any 

color space that can be described by a matrix profile (do not 

use LUT profiles!). It is mandatory that every image here has 

the respective profile embedded. You can replicate any image 

of the main text in the color appendix, and also add further 

images (at the discretion of the Editorial and Production 

Staff). You should add a description of how the images have 

been prepared, and what the reader should do in order to 

achieve an appropriate reproduction, i.e., specify the rendering 

intent. If the color space is large, an appropriate reproduction 

may be possible only on a monitor, or on a specific 

type of large gamut printer, etc. Let the readers know this. 

Readers’ Guidelines 

All color images in the main body of every paper have been 

prepared for sRGB. 

Monitor Viewing of Color Articles 

The best way to enable faithful viewing of color images in 

articles is a calibrated display with a properly installed monitor 

profile. If you possess a monitor calibrator, make sure 

that the monitor is properly calibrated and profiled. If not, 

and you have a Mac, nothing else is required; just make sure 

that the monitor settings are somewhat reasonable (white 

does not look pink etc.). If you don’t have a monitor calibrator 

and you have a PC, follow the guidelines in the section 

above, “Simple Guidelines for Achieving sRGB-like behavior 

on PCs.” 

On the Macintosh, most known PDF viewing software 

(including Preview and Acrobat) obeys embedded color profiles. 

In Windows, Acrobat also supports profiles, and makes 

use of monitor profiles, but defaults to sRGB if none has 

been installed on the system. 

Printing of Color Articles 

Printing color is less predictable than displaying color on a 

monitor. Depending on the printer type, paper, inks, printing 

speed, driver settings, etc., quite different results may 

occur. Though printer profiling can correct deviate color 

behavior, it cannot perform magic given sometimes very 

limited color gamuts. Hence, if color quality is essential to an 

article, make sure that you use a good printer with reasonable 

inks, quality coated paper, and high-quality driver 

settings. 

If you know how to profile a printer, make use of this 

capability. Otherwise we recommend using the default settings 

of the printer driver, which should lead to reasonable 

results, since most printer manufacturers assume, more or 

less, an sRGB color space for the source data, i.e., the color 

space for which the article images have been prepared. 

If the author(s) have included a color plate appendix 

with specially prepared images (larger color gamut etc.) they 

should also have provided special guidelines to get to the 

desired result, e.g., by specifying the rendering intent. These 

color plates are usually intended for experts who should 

strictly follow the authors’ directives, and employ color 

management. 

Appendix 

To check if your system obeys color profiles, the following 

site provides a test document with a simple test to see if a 

PDF viewer supports embedded profiles: http:// 

www.color.org/version4ready.html. 

—Patrick Herzog 

X-Rite, Inc. 

v

IS&T Corporate Members 

IS&T Corporate Members provide significant financial support, thereby assisting the Society in achieving its goals of 

disseminating information and providing professional services to imaging scientists and engineers. In turn, the Society 

provides a number of material benefits to its Corporate Members. For complete information on the Corporate Membership 

program, contact IS&T at info@imaging.org. 

Sustaining Corporate Members 

Adobe Systems Inc. 

345 Park Avenue 

San Jose, CA 95110-2704 

Canon USA Inc. 

One Canon Plaza, Lake Success 

New York, NY 11042-1198 


343 State Street 

Rochester, NY 14650 

Hewlett-Packard Company 

1501 Page Mill Road 

Palo Alto, CA 94304 

Supporting Corporate Members 

Lexmark International, Inc. 

740 New Circle Road NW 

Lexington, KY 40511 

Xerox Corporation 

Wilson Center for Research and 

Technology 

800 Phillips Road 

Webster, NY 14580 

Fuji Photo Film Company, Ltd. 

210 Nakanuma, Minami-ashigara 

Kanagawa 250-0193 Japan 

Konica Minolta Holdings Inc. 

No. 1 Sakura-machi 

Hino-shi, Tokyo 191-8511 Japan 

Donor Corporate Members 

ABBY USA Software House, Inc. 

47221 Fremont Blvd. 

Fremont, CA 94538 

ILFORD Imaging Switzerland GmbH 

Route de l’Ancienne Papeterie 1 

CH-1723 Marly, Switzerland 

Axis Communications AB 

Embdalavägen 14 

SE-223 69 Lund, Sweden 

Cheran Digital Imaging & Consulting, Inc. 

798 Burnt Gin Road 

Gaffney, SC 29340 

Clariant Produkte GmbH 

Division Pigments & Additives 

65926 Frankfurt am Main Germany 

Felix Schoeller Jr. GmbH & Co. KG 

Postfach 3667 

D-49026 Osnabruck, Germany 

Ferrania SpA 

Viale Martiri Della Liberta’ 57 

Ferrania (Savona) I-17014 

GretagMacbeth 

Logo GmbH & Co. KG 

Westfälischer Hof Garbrock 4 

48565 Steinfurt, Germany 

Hallmark Cards, Inc. 

Chemistry R & D 

2501 McGee, #359 

Kansas City, MO 64141-6580 

MediaTek Inc. 

No. 1 Dusing Rd., 1 

Hsinchu 300 R.O.C, Taiwan 

Nitta Gelatin NA Inc. 

201 W. Passaic Street 

Rochelle Park, NJ 07662-3100 

Pantone, Inc. 

590 Commerce Blvd. 

Carlstadt, NJ 07072-3098 

Quality Engineering Associates (QEA), 

Inc. 

99 South Bedford Street, #4 

Burlington, MA 01803 

The Ricoh Company, Ltd. 

16-1 Shinei-cho, Tsuzuki-ku 

Yokohama 224-0035 Japan 

Sharp Corporation 

492 Minosho-cho, Yamatokoriyama 

Nara 639-1186 Japan 

Sony Corporation/ 

Sony Research Center 

6-7-35 Kita-shinagawa 

Shinagawa, Tokyo 141 Japan 

as 12/3/06

Journal of Imaging Science and Technology® 51(1): 1–22, 2007. 

© Society for Imaging Science and Technology 2007 

Improved Pen Alignment for Bidirectional Printing 

Edgar Bernal and Jan P. Allebach 

School of Electrical and Computer Engineering, Purdue University, West Lafayette, IN 47907-1285 

E-mail: eabernal@purdue.edu 

Zygmunt Pizlo 

Department of Psychological Sciences, Purdue University, West Lafayette, IN 47907-1285 

Abstract. The quality of the prints produced by an ink jet printer is 

highly dependent on the characteristics of the dots produced by the 

ink jet pens. While some literature discusses metrics for the objective 

evaluation of print quality, few of the efforts have combined 

automated quality tests with subjective assessment. The authors 

develop an algorithm for analyzing printed dots and study the effect 

of the dot characteristics on perceived print alignment. The authors 

establish the perceptual preferences of human observers via a set 

of psychophysical experiments. 

© 2007 Society for Imaging Science and Technology. 

DOI: 10.2352/J.ImagingSci.Technol.200751:11 

INTRODUCTION 

The advent of low cost, photo-quality ink jet printers has 

raised the need for an objective means of determining print 

quality that is consistent with what the end-user perceives. 

High level quality metrics have been specified in the International 

Association for Standardization/International Electrotechnical 

Commission (ISO/IEC) guidelines on hardcopy 

print assessment. 1 These guidelines include metrics for four 

distinct categories of printed areas: line and character metrics, 

solid fill metrics, tint solid metrics, and background 

field metrics. Other metrics that enable the quantification of 

performance aspects relevant to ink jet printers have also 

been proposed. These aspects include color registration, 

color consistency, modulation transfer function (MTF), text 

quality, sharpness, 2 dot quality, and line quality. 3,4 

Multiple efforts have been made to automate the process 

of image quality assessment both during product 

development 5 and manufacturing, 6 and for benchmarking 

and competitive analysis. 4,7 The ultimate objective of these 

initiatives is to provide the ability to measure a large volume 

of prints and, at the same time, achieve the repeatability and 

objectivity that visual inspection-based processes lack. 

Attempts have also been made to characterize and reduce 

print quality defects inherent to ink jet technology, 

such as the inability to achieve uniformity in areas of solid 

color because of banding, 8 printing artifacts derived from 

incorrect dot placement, 9 dot shapes and sizes that differ 

from the ideal, 10 and the presence of tails and satellites due 

to aerodynamic effects. 11 

Received Jul. 31, 2006; accepted for publication Oct. 11, 2006. 

1062-3701/2007/511/1/22/$20.00. 

Low level models have also been used to improve the 

quality of printed halftone images. There are two approaches 

to the development of such model-based algorithms. 12 The 

first approach uses models that reflect the actual process 

whereby the digital halftone is transformed to colorant on 

the page. For example, models for the laser beam, exposure 

of the organic photoconductor, and the resulting absorptance 

on the paper have been embedded into the Direct 

Binary Search (DBS) halftoning algorithm for electrophotographic 

(EP) printers, 13 showing good improvement over 

regular binary DBS with tone correction. The second approach 

is largely based on characterization of the halftone 

image as it exists on the printed page. For example, analytical 

and stochastic models for EP printer dot interactions 

have been incorporated in the DBS halftoning algorithm, 14 

yielding enhanced detail rendition and improved tonal gradation 

in shadow areas. For ink jet printers, the displacement 

and profile of individual dots were measured and the 

conditional pixel statistics were calculated. 15 These results 

were then applied to the DBS halftoning algorithm to develop 

an ink jet printer model that reduced the visual 

artifacts caused by systematic and random errors in dot 

placement. 

An ink jet printer places marks on the page by means of 

a print head that contains columns of nozzles through which 

ink is fired. The nozzles are fired in a carefully controlled 

manner as the print head moves back and forth across the 

page. Careful alignment of the dot patterns printed in successive 

passes across the page is critical to perceived print 

quality. The aim of this paper is to study the effects of the 

printed dot characteristics on the perception of ink jet pen 

alignment via an approach that relies both on automated 

image analysis tools and psychophysical experiments. We develop 

a set of image analysis tools to characterize many attributes 

of printed dots, including alignment. We also examine 

the relationship between physical alignment and 

perceived alignment. This paper focuses on the HP DeskJet 

6540 (Hewlett-Packard Company, 3000 Hanover St., Palo 

Alto, CA 94304-1185) high resolution ink jet printer with 

plain paper, but the methodology is generally applicable to 

other ink jet printers and paper types as well. 

The structure of the paper is as follows: we first give an 

overview of the ink jet printing process. We then describe 

the calibration of the image capture device and the design of 

1

Bernal, Allebach, and Pizlo: Improved pen alignment for bidirectional printing 

the tools that enable alignment measurement and dot analysis. 

We present some experimental results obtained from the 

application of the dot analysis tool to test prints. We proceed 

to describe the set of psychophysical experiments that were 

performed on alignment perception. Finally, we give our 

conclusions. 

PRELIMINARIES 

Figure 1 illustrates the operation of a typical ink jet printer. 

The paper is advanced through the unit by a series of rollers 

driven by a stepper motor. A carriage transports the pen or 

printhead back and forth across the page. The printhead 

consists of one or more columns of nozzles through which 

drops of ink are fired onto the surface of the paper. Printed 

dots reveal artifacts that depend on print options such as 

print resolution, speed, directionality, and the number of 

printing passes over each pixel on the paper. A print mode 

specifies the set of such print options with which a document 

is printed. The pixels that are printed in a given pass 

across the page comprise a subset of the pixels in a horizontal 

band with height equal to the height of the print head. 

This horizontal band of pixels is called a swath. In the 

single-pass print modes, the printhead passes only once over 

each position on the paper, so the swaths do not overlap. For 

a multipass print mode with N passes, the paper only advances 

a fraction 1/N of the height of the printhead between 

passes. With the single pass print modes, misalignment between 

adjoining swaths is especially visible. With multipass 

modes, the misalignment is masked to some extent by the 

overlapping swaths. Typically, a print mode with one pass, a 

higher printhead velocity, and lower resolution is used for 

draft quality printing and a mode with multiple passes, a 

lower printhead velocity, and a higher resolution is used for 

the highest quality printing. To achieve print resolutions that 

are lower than the native resolution of the print mechanism, 

two or more dots are printed in a cluster for each pixel. 

In this paper, we are primarily interested in draft quality 

printing of black and white documents using a single pass 

mode. This modus operandi implies that there is a tradeoff 

between print speed and print resolution. To see this, consider 

the simpler case in which the printhead has only one 

column of nozzles and is moving at a speed of v inches 

per second (ips) across the page. Suppose also that the 

maximum frequency at which the nozzles can be fired is f 

firings/sec. Then, the closest distance at which two horizontally 

adjacent dots can be printed is d=v/f in., and the 

maximum resolution that can be achieved with that particular 

print mode is 1/d dots per inch (dpi). Since f is fixed for 

a given printhead, the print resolution is inversely proportional 

to the print speed. In unidirectional print modes, the 

pen only fires ink while it is traveling in one direction across 

the page (either while traveling from left to right or from 

right to left), while in bidirectional print modes, successive 

swaths are printed in opposite directions. 

When printing at a resolution of 300 dpi, the DeskJet 

6540, which has a pen with vertical nozzle-to-nozzle spacing 

of 1/600 in., renders a single dot as two vertically adjacent 

dots. However, given the high nozzle firing frequency required 

to print at high carriage speeds, some of the nozzles 

fail to fire ink occasionally, which results in some single dots 

being printed on the page. Figure 2 shows typical single and 

double dots printed at a carriage speed of 30 ips and 

scanned at 7000 dpi with a QEA IAS-1000 Automated Image 

Analysis System (Quality Engineering Associates Inc, 25 

Adams Street, Burlington, MA 01803). 

Figure 3 shows the appearance of a typical dot printed 

with a single-pass, 300 dpi resolution print mode with different 

carriage speeds and printing directions. It illustrates 

the fact that as print speed increases, the dot shape becomes 

more asymmetric, and thus more dependent on the printing 

direction. Other artifacts that are related to print speed are 

Figure 1. Operation of an ink jet printer: a the 3-D view illustrates the 

movement of the printhead and b the cross-section illustrates the paper 

path. 

Figure 2. Effect of print resolution on dot appearance: a single dot and 

b double dot printed with 300 dots per inch dpi, 30 inches per second 

ips, right-to-left print mode. Scanned at 7000 dpi with QEA 

IAS-1000. 

2 J. Imaging Sci. Technol. 511/Jan.-Feb. 2007


Figure 4. Other artifacts due to high print speeds: a satellites and b 

tails on dots printed at 300 dpi, 60 ips, right-to-left print mode. Scanned 

at 7000 dpi with QEA IAS-1000. 

Figure 3. Typical dot printed at 300 dpi: a 15 ips left-to-right print 

mode, b 45 ips left-to-right print mode, and c 45 ips right-to-left print 

mode. Scanned at 7000 dpi with QEA IAS-1000. 

tails and satellites, which occur when the drop of ink breaks 

up as it exits the print nozzle. If the secondary droplet breaks 

away completely from the main droplet, it forms a satellite 

[see Fig. 4(a)], and if it breaks away only partially, it forms a 

tail [see Fig. 4(b)]. Tails and satellites usually trail the main 

dot relative to the direction of travel of the pen. Since there 

is a tradeoff between print quality and print speed and also 

because the media characteristics and page content impact 

the choice of print mode that will yield the best print quality, 

a number of different print modes are typically designed for 

an ink jet printer. The specific effect of the print modes on 

the dot attributes will be described in detail later. 

The process of printing a vertical line with a single-pass, 

bidirectional mode is illustrated in Fig. 5 for a simplified 

printer architecture. The printhead contains nozzles (in this 

case, 3 columns of 8 nozzles each) that fire the colorant onto 

the page. Typically, a real printhead would contain many 

more nozzles. For example, the black ink printhead for the 

HP DeskJet 6540 printer contains 4 columns of 168 nozzles 

each. The two-dimensional image of the line (including the 

blank regions surrounding the line) is encoded onto a print 

mask, 16 which consists of a two-dimensional array of 0’s and 

1’s. A 1 indicates firing the nozzle at that particular position 

and a 0 indicates no firing. In the case illustrated by Fig. 5, 

the upper segment of the vertical line is printed on the leftto-right 

pass of the pen and the lower segment is printed on 

the right-to-left pass. The size of each swath is determined 

by the distance between the top and bottom nozzles in the 

pen. 

Vertical alignment within a swath is readily achieved via 

the fixed spatial positions of the nozzles in the printhead, 

and between swaths by the correct advancement of the paper. 

Horizontal alignment within a swath is also readily 

achieved by virtue of the fixed spatial configuration of the 

nozzles in the print head, and through synchronized firing of 

J. Imaging Sci. Technol. 511/Jan.-Feb. 2007 3


ink jet printer and scanned with the Aztek Premier high 

resolution drum scanner (Aztek Digital Imaging, 13765-F 

Alton Parkway, Irvine, CA 92618). We developed a software 

tool that classifies and quantifies the printed dot characteristics 

and calculates the relative position of adjacent swaths 

from the scanned version of the test pattern. To match the 

perceived attributes and the measured quantities, we used a 

set of test pages encoded in a low level printing language as 

the stimuli in the psychophysical experiments. The low level 

printing language allows fine tuning of the swath-to-swath 

offsets as well as print speeds and print directions. 

Figure 5. Illustration of the process of printing a vertical line in a singlepass, 

bidirectional print mode, with a 24-nozzle pen. The vertical position 

of the pen with respect to the media changes from swath to swath as the 

paper is advanced. 

the nozzles while the print head moves at constant velocity. 

Between swaths, horizontal alignment depends on the timing 

of the start of the firing of nozzles at the initial edge of 

the page. Consequently, swath-to-swath horizontal alignment 

is the factor that ultimately determines whether or not 

the print appears aligned to the viewer. Figure 5 illustrates 

the situation where an undesired line break is produced due 

to inaccurate horizontal alignment between swaths. In reality, 

however, the line segments printed on each of the swaths 

are more complex than those depicted in the figure. This is 

because of the dot irregularities and the fact that the relationship 

between the main dot and tails or satellites is reversed 

from raster to raster. Thus, the task of achieving accurate 

swath-to-swath alignment requires knowledge of how 

the human viewer actually perceives the position of the main 

dot/satellite or main dot/tail pair. 

The ability of the human viewer to detect misalignment 

has been widely studied in cases where the line segments are 

displayed or printed with ideal devices. The just noticeable 

angular offset between two line segments is called Vernier 

acuity. 17 It has been found that the discriminable offset 

ranges from 5 to 10 seconds of arc (2.910 −4 in. to 5.8 

10 −4 in. at a viewing distance of 12 in.), which is much 

less than the distance of 25 seconds of arc between foveal 

receptors. However, few studies have considered the case 

where the lines are composed of irregular dots. Patel et al. 

found that thresholds for asymmetric irregular shapes were 

higher than those for regular dots. 18 Since dots become more 

irregular as the print speed increases, evaluation of alignment 

perception at high print speeds (45 ips and above) is 

of particular interest. Also, since higher print speeds imply 

lower print resolutions, the test resolution was fixed at 

300 dpi for the fastest print modes. This is the highest resolution 

achievable at the highest print speed for this printer. 

To enable automatic measurement of the print characteristics, 

we designed a test pattern that is printed with an 

PREPROCESSING 

The alignment measurement procedure consists of printing, 

scanning, and processing a test pattern in order to get dot 

placement information. Even though the images obtained 

with the QEA System are sharper than those obtained with 

the Aztek Scanner, the latter was chosen for this task due to 

its larger field of view at high resolutions. The alignment 

analysis tool relies on averaging dot positions across a large 

number of dots that cover a printed area of approximately 

1 in.1 in. The Aztek Scanner is capable of capturing a 

region of 8.5 in.11 in. regardless of the scanning resolution, 

while the field of view of the QEA is less than 0.1 in. 

0.1 in. at 8000 dpi. In this section, the scanner calibration 

procedure that allows the mapping of the scanner grayscale 

output into absorptance is described. Also, the design of the 

test pattern and the initial processing to find boundaries 

between dots are presented. 

Scanner Calibration 

Scanner calibration is the process whereby device-dependent 

scanner RGB values are converted into values of a deviceindependent 

color space such as CIE XYZ. 19 The scanner 

calibration was performed as suggested in Ref. 20: 

1. A TIFF file containing 17 half-inch square test 

patches with gray values ranging from 0 to 1 was 

generated. 

2. The TIFF file was printed using the printer driver’s 

halftoning technique at 600 dpi. The same printer 

and the same colorant (K) used in the alignment 

study were used in the calibration process. 

3. The luminance values of the patches were measured 

with a calibrated Gretag SPM-50 (Gretag Data and 

Image Systems, Althardstrasse 70, CH-8105 Regensdorf, 

Zürich, Switzerland) spectrophotometer. Five 

measurements were taken for each patch and the results 

were averaged. The resulting luminance was 

converted to absorptance (0–1) values and then rescaled 

to fall in the range 0–255. 

4. The patches were scanned at 1000 dpi with the Aztek 

Premier drum scanner. The resulting patch images 

were cropped to avoid edge effects, and the average 

grayscale value of each patch was found. 

5. The scanner data S was fitted to the spectrophotometer 

data G using an exponential function of the 

form G=a 1 S/255 +a 2 by minimizing the mean- 



squared error between the function output and the 

data points. The resulting coefficients were 

a 1 =262.48, =1.23, and a 2 =5.02. 

6. Before any scanned image is processed, it is calibrated 

using this mapping. The raw data and the 

fitted curve are shown in Fig. 6. 

Test Pattern Design and Dot Boundary Calculation 

The first step toward pen characterization consists of designing 

a test pattern with attributes that enable the measurement 

of the quantities of interest. In our case, we are interested 

in being able to measure swath-to-swath alignment 

and to quantify dot characteristics such as shape, size, elongation, 

and presence or absence of artifacts, such as tails and 

satellites. 

The test pattern we designed is a 600600 pixel grid 

where only every 20th row and 20th column contains a 

printed dot. Hence, there are a total of 900 dots in the 

printed test pattern. In order to facilitate scanner focusing 

and to stabilize the pen’s nozzle firing, a 50-pixel-wide solid 

frame surrounds the central grid, and a 400 pixel 

400 pixel solid black region is placed on each side of the 

frame. Figure 7 shows the designed test pattern. 

The test pattern is printed in the desired print mode 

(the dot analysis tool works for any print mode, as long as 

the test pattern complies with the specifications listed above) 

and then scanned at a resolution of 8000 dpi with the Aztek 

Premier Scanner. The scanned image is processed to produce 

a binary segmentation mask image that indicates the presence 

or absence of ink at every pixel. The threshold for the 

image binarization is calculated according to Otsu’s 

method, 21 an unsupervised approach that minimizes the 

intra-class variance of the black and white pixels. Figure 8 

shows a portion of the scanned test pattern and its corresponding 

segmentation mask. 

With the aid of the segmentation mask, boundaries between 

rows and columns are found, and boundaries delimiting 

dot regions are determined. Boundaries between columns 

are determined by vertically projecting the data of the 

binary image and finding the points of the projection that 

are greater than zero, as illustrated in Fig. 9(a). The process 

is similar for row boundaries, except that the projection is 

done horizontally. The boundaries for a dot’s cell are determined 

by intersecting the boundaries of the row and the 

column to which the dot belongs, as illustrated in Fig. 9(b). 

The centroid of each dot is then calculated based on the 

spatial distribution of ink absorptance throughout the dot’s 

corresponding cell. If the cell of the dot is defined by the 

coordinates x 1 ,y 1 and x M ,y N , as shown in Fig. 9(b), then 

its horizontal center of mass is given by 

C x = 

N M 

 

n=1 m=1 

Im,nx m 

N M 

 

n=1 m=1 

Im,n 

, 1 

where Im,n is the absorptance value of the image at the 

pixel with coordinates x m ,y n . Similarly, the vertical center 

of mass is given by 

C y = 

N M 

 

n=1 m=1 

Im,ny n 

N M 

 

n=1 m=1 

Im,n 

. 2 

Figure 6. Raw data and fitted curve for the Aztek Premier Scanner. 

Figure 7. Test pattern for printhead and alignment characterization. 

Figure 8. Cropped version of a test pattern printed with 15 ips, bidirectional 

print mode and scanned at 8000 dpi with Aztek Premier Scanner 

and b corresponding binary mask. 



DOT ANALYSIS 

In this section, the procedure for misalignment measurement, 

dot analysis, and pen characterization is presented. 

First, we will describe the procedure for measuring misalignment 

from scanned images of the test target. Then, we will 

discuss the algorithms that classify dots into double and 

single dots, segment double dots, and detect tails and satellites 

and separate them from the main dots. These algorithms 

were applied to images obtained with the Aztek 

Scanner. 

Misalignment Measurement 

Since the height of each swath is known, it is possible to 

determine the regions in the image that correspond to different 

swaths by segmenting the image file into horizontal 

stripes with height equal to the height of one swath. Then, if 

the upper and lower halves of the test pattern shown in Fig. 

7 are positioned in adjacent stripes, misalignment can be 

estimated by calculating the offset between the average horizontal 

position of the dots in the upper half of the pattern 

and the average horizontal position of the dots in the lower 

half of the pattern. If C xi,j is the horizontal center of mass of 

the dot in the ith row and jth column, then the average 

swath-to-swath misalignment is given by 

C x = 1 

450 j=1 

 

30 15 

C xi,j − C xi+15,j 

i=1 

because rows 1 to 15 belong to the upper swath and rows 16 

to 30 belong to the lower swath, and there are a total of 450 

dots in each swath. This approach, however, yields estimates 

that are highly dependent on the image skew, which can 

occur during both printing and scanning. 

In order to account for the effect of image skew, the 

angle of skew must be estimated. This is done by fitting a 

straight line to each of the rows of dot centroids via orthogonal 

regression 22 and averaging the slopes of the set of 

straight lines thus obtained. The new reference columns are 

found by fitting straight lines to each of the columns of dot 

centroids, with the constraint that they should be perpendicular 

to the line describing the skew of the image. The 

orthogonal distance of each of the centroids to its respective 

reference column is calculated. The average of these distances 

across dots on each swath is computed to find the 

average offset of each swath. The total misalignment is estimated 

by computing the difference between the average offset 

of the upper swath and the average offset of the lower 

swath. Figure 10 illustrates the process of skew estimation 

and misalignment measurement. 

Dot Classification 

As seen earlier, double dots are inherent to 300 dpi resolution 

print modes when printing with a 600 dpi resolution 

3 

Figure 9. a Finding boundaries between rows and columns and b 

finding the centroid of a dot. 

Figure 10. Skew estimation and misalignment measurement. 



printhead. Also, as the print speed increases, tails and satellites 

appear more frequently. Identifcation of the main attributes 

of the printed dots plays a fundamental role in the 

dot analysis process. The process of dot classification into 

single and double dots consists of coding the most relevant 

information of the dot image and comparing it to a database 

of previously coded training samples to find the one that 

most resembles the dot. To this end, the principal components 

of the distribution of the information embedded in 

the set of training dot images must be found. 23 

The simplest approach consists of representing the 

NN image of the dot as an N 2 1 vector in an 

N 2 -dimensional space. Then, if the set of training samples 

consists of the images I 1 ,I 2 ,...,I M , we can represent each 

image I i asavector i . The average image is given by 

M 

= 1 i . 

M i=1 

The principal components of the set of training images are 

the eigenvectors of the covariance matrix 

M 

C = 1 i − i − T . 

M i=1 

This set of vectors is the basis of the new feature space. 

Let v 1 ,v 2 ,...,v K denote the set of K eigenvectors corresponding 

to the K largest eigenvalues of C. This set will be 

the basis of the new eigenspace and any NN 

arbitrary dot can be approximated by a linear 

combination of its elements as K 

i=1 i v i +, where 

i =v T i −. Since the basis of the space is fixed, an image 

− can be represented by the vector of its coefficients, 

= 1¯ K . The training of the algorithm consists of calculating 

the coefficients 1 , 2 ,..., M that correspond to 

the images 1 , 2 ,..., M whose class is known. To classify 

4 

5 

anewdot, its corresponding coefficients are found and 

the Euclidean distance i =− i is calculated for 

i=1,...,M. The new dot is assigned to the same class as dot 

j, where 

j = argmin i ,i =1,2, ...,M, 

i 

i.e., we find the dot j from the training set that is closest to 

the new dot in terms of the K coefficients and assign the new 

dot to the same class to which dot j belongs. 24 In our case, 

the training set consisted of five single dots and five double 

dots, and the classification stage worked with four coefficients, 

which implies that M10 and K4. Figure 11(a) 

shows a sample image that illustrates the results of the dot 

classification stage. Dots surrounded by a single frame were 

identified as single dots and dots surrounded by a double 

frame were identified as double dots. The performance of 

the classification stage was found to be 100% accurate 

among the group of patterns tested. This group was comprised 

of at least 100 test patterns, each composed of 900 

dots. Figure 11(b) shows a scatter diagram of the coefficients 

1 and 2 for the single and double dot training samples 

and for the single and double dots in Fig. 11(a). It can be 

seen that in this two-dimensional feature space, the projection 

coefficients form two clusters, one corresponding to 

each dot class. This is why a simple metric such as the 

Euclidean distance yields a good classification performance. 

Dot Bisection 

All dots identified as double dots have to go through the 

process of bisection. This is necessary because in the end we 

want to know the characteristics of individual dots. Given 

the large number of dots present in a single test pattern, 

there is a need to implement an efficient segmentation algorithm. 

Caselles et al. 25 and Kass et al. 26 devised segmentation 

algorithms based on active contours that lock onto image 

6 

Figure 11. Operation of the dot classification stage: a Cropped region of a test image after the dot 

classification stage. Dots surrounded by a single frame were identified as single dots and dots surrounded by 

a double frame were identified as double dots. b Scatter diagram of coefficients 1 and 2 for the training 

samples and for the dots in Fig. 11a. 



features such as lines and edges. A priori knowledge of the 

topology of the desired final solution imposes an important 

constraint on the possible approaches and allows for the 

design of a faster algorithm than those belonging to the 

active contour class, which were designed for a general class 

of images. 

Our solution is a fixed-topology, multi-resolution approach 

to the “snakes” active contour model proposed by 

Kass et al. 26 This model modifies the shape of the solution 

until a contour with minimum total energy is found. Let the 

contour be described parametrically as vs=xs,ys, s 

0,1. Lets i N i=1 be a set of real numbers such that 0 

s 1 ¯ s N 1. Then the total energy of the contour can 

be approximated by 

E = E cont s i + E curv s i + E image s i , 

i 

where E cont is the energy due to the continuity of the contour 

components, is its corresponding scaling factor, E curv 

7 

is the energy due to curvature or bending of the contour, 

is its corresponding scaling factor, E image is the energy due to 

the image gradient on the contour components, and is its 

corresponding scaling factor. Minimizing the continuity energy 

corresponds to finding a contour in which the distance 

between elements is small. Minimizing the curvature energy 

is equivalent to finding a contour with the smallest curvature 

possible. Lastly, minimizing the image energy corresponds to 

finding a contour with elements located in small gradient 

image regions. Minimizing the overall energy corresponds to 

finding a compromise between the three energy values regulated 

by the three scaling constants. 

As will be seen later, neither the continuity term (which 

regularizes the interpixel distances) nor the curvature term 

(which controls the smoothness of the contour) imposed in 

the snakes approach were utilized herein, since they are implicit 

in the implementation of our algorithm. For the external 

energy term, image absorptance rather than image gradient, 

as suggested by Kass et al., 26 was chosen. This is 

Figure 12. First stage of bisection process: a candidates for endpoints of the bisecting contour, b selected 

endpoints, c candidates for third component of the contour, and d selected point. 



because the bisecting contour should lie in the lightest path 

between the two dark regions that correspond to each of the 

dots. 

Figure 12 illustrates the evolution of the dot bisection 

process. In the first stage of the process, the line segment 

with the lowest integrated absorptance per unit length between 

the left boundary of the dot image and the dot centroid 

is found. The leftmost vertex of the bisecting contour is 

the one at which this particular segment originates. The line 

segment with the lowest integrated absorptance per unit 

length between the dot centroid and the right boundary of 

the dot image is also found. The rightmost vertex of the 

bisecting contour is the one at which this particular segment 

terminates. This step is illustrated in Fig. 12(a), which shows 

the set of candidates for contour endpoints and their respective 

line segments. Figure 12(b) shows the selected vertices. 

The third vertex of the contour is the point equidistant to 

the endpoints such that the line segments between it and the 

endpoints have the least integrated absorptance per unit 

length. Figure 12(c) shows some of the points in the set of 

candidates for a new vertex of the contour and the corresponding 

line segments. The average absorptance per unit 

length between each candidate point and both of the endpoints 

of the contour is calculated. The point that defines 

the segments with the least integrated absorptance per unit 

length is kept, as shown in Fig. 12(d). Note that the candidate 

points for new vertices are all located on the line segment 

that lies halfway between the endpoints and which is 

perpendicular to the line connecting the two endpoints. 

Thus, they are all equidistant to both endpoints. Also, note 

that the search is limited to a specific angle, called the angle 

of sweep. In the case illustrated in Fig. 12, the angle of sweep 

was set to ±15°. 

In the subsequent stages of the algorithm, the same procedure 

is implemented between intermediate contour vertices. 

The multiresolution effect is a natural consequence of 

the fact that as the procedure advances, the energy of the 

active contour is minimized on smaller regions of the image. 

The distance between the contour vertices is determined by 

the number of stages in the procedure: the higher the num- 

Figure 13. Evolution of bisection process after stages a 1, b 2, c 3, and d 4. 



Figure 15. Initialization of tail detection algorithm: a fitted ellipse for 

single dot with a tail and direction of projections orthogonal to major axis 

of ellipse and b projected absorptance profile and positions of center of 

mass and local minimum in profile. 

Figure 14. Illustration of ellipse fitting a grayscale image of dot, b 

binary dot, and c dot outline and fitted ellipse. 

ber of stages, the higher the number of vertices in the contour, 

and thus the smaller the distance between vertices. The 

curvature characteristics of the contour are determined by 

the magnitude of the angle of sweep illustrated in Fig. 12(a). 

The larger the magnitude of that angle, the less smooth the 

contour can be. Thus, the continuity and curvature constraints 

are implicit in the implementation of the algorithm. 

Figures 13(a)–13(d) illustrate the evolution of the process. 

Ellipse Fitting 

Ellipse fitting is a basic task in pattern recognition because it 

describes the data in terms of a geometric primitive, thus 

reducing and simplifying its representation. In our case, ellipse 

fitting is used to estimate dot eccentricity, aspect ratio, 

and orientation. Historically, techniques for ellipse fitting are 

divided into two main approaches: clustering 27,28 and leastsquares 

fitting. 29,30 While clustering methods are robust to 

outliers and can detect multiple primitives at once, they are 

computationally expensive and have low accuracy. On the 



Table I. Information provided for each single dot. 

Output 

Location of dot’s center of mass 

Total integrated absorptance of the dot 

Coefficients of the ellipse fitted to the dot 

outline 

Information of whether dot has a tail or not 

If dot has a tail, the location of the components 

of the tail-segmenting contour 

If dot has a tail, the location of the main dot 

and tail’s centers of mass 

If dot has a tail, the total integrated absorptance 

of the main dot and of the tail 

Format 

21 vector of double precision 

floating point numbers 

Double precision floating 

point number 



Binary number 

29 array of double precision 


22 array of double precision 




Figure 16. Tail detection algorithm: a single dot with a tail and b tail 

segment. 

other hand, the least-squares methods are fast and accurate, 

but can only fit one geometric shape at a time and are more 

sensitive to outliers. 30 We found that the model proposed by 

Halif and Flusser, 30 which is an improved version of that 

proposed by Fitzgibbon et al., 29 performed accurately and 

efficiently enough for our purposes. A short description of 

the method can be found in the Appendix (Appendix available 

as Supplemental Material on the IS&T website, 

www.imaging.org). The ellipse is fitted to the set of coordinates 

of the pixels that belong to the dot outline defined by 

the binary image of the dot, as shown in Fig. 14. The binary 

image of the dot is obtained by thresholding its grayscale 

image in the same manner as the binary segmentation mask 

is obtained from the grayscale scanned image (as described 

in section entitled “Test Pattern Design and Boundary Calculation”). 

From the ellipse coefficients, quantities such as 

dot aspect ratio and orientation are estimated. 

Tail Detection 

Tail and satellite dots manifest themselves in a way very similar 

to that in which double dots appear on the printed page: 

there is a region of low absorptance between two regions of 

higher absorptance. In the case of double dots, these regions 

correspond to the two main dots, while in the tail/satellite 

problem, they correspond to the main dot and the tail or 

satellite. The main difference between the two is the fact that 

the direction of the segmenting contour that separates the 

tail or the satellite from the main dot is perpendicular to the 

orientation of the dot. From the ellipse-fitting stage, we can 

estimate the orientation of the main dot by the inclination of 

the main axis of the ellipse that best fits the points on the 

outline of the dot. 

Figure 15(a) shows a dot with a tail and its fitted ellipse. 

An absorptance profile is obtained by projecting the dot absorptance 

in the direction perpendicular to the ellipse orientation, 

as indicated by the arrows. Figure 15(b) shows the 

profile obtained for this particular dot. For instance, the projected 

absorptance value corresponding to the path highlighted 

by the dashed gray (black) arrow in Fig. 15(a) is the 

point in the profile of Fig. 15(b) marked with the dashed 

gray (black) line. Starting at the point in the profile that 

corresponds to the dot’s center of mass [see dashed gray line 

in Fig. 15(b)], a search for a local minumum is performed 

[see dashed black line in Fig. 15(b)]. The existence of a local 

minimum in the profile indicates the presence of a tail. If 

there is at least one local minimum, the position of the local 

minimum closest to the center of mass is found. In order to 

decrease the false alarm rate in the tail detection process, the 

decision that a tail is present is made only if the value of the 

profile at this local minimum is at least 20% smaller than the 

maximum value of the profile. 

The tail-segmenting contour is initialized at the extreme 

points of the line segment whose projection yielded that 

particular local minimum [in this case, the segmenting contour 

is initialized at the end points of the dashed black arrow 

in Fig. 15(a)]. The subsequent stages of the tail separation 



process are the same as in the dot bisection process: at each 

stage, the point equidistant to the endpoints, such that the 

line segments between it and the endpoints have the least 

integrated absorptance per unit length, is found and added 

to the contour. This strategy makes the overall procedure for 

separating the main dot and its tail robust to errors in the 

initial estimation of the local minimum based on the projected 

absorptance. Figure 16 shows the results of the tail 

detection algorithm applied to a single dot with a tail. 

EXPERIMENTAL RESULTS FOR DOT SHAPE 

ANALYSIS 

In this section, some of the results gathered from the application 

of the dot analysis tool to different test pages are 

presented. The objective of the tests was to establish the 

variability of the dot characteristics from pen to pen for a 

sample population of pens, and from print mode to print 

mode for a single pen. 

Output of Dot Analysis Tool 

Pen alignment has an important impact on print quality, and 

the precision with which alignment is controlled impacts 

product engineering and cost. Dot shape characteristics impact 

both the appearance of the printed page and the way 

alignment is perceived. Therefore, in order to thoroughly 

study how alignment is perceived by human viewers, we 

must first understand how dot shape characteristics vary 

with the print mode for a single pen. However, these results 

will only be meaningful if we first establish that printing 

properties across a population of pens for a given print 

mode remain more or less stable. Thus, we first examine this 

aspect of the pen characteristics. 

The dot analysis tool takes the scanned image of the test 

pattern (see Fig. 7) printed with the HP DeskJet 6540 and 

processes it in the manner described in the preceding section. 

The output of the analysis tool is a set of text files that 

contain all the information required to extract the characteristics 

of each dot in the printed pattern. For each dot, the 

information of whether it is single or double is provided. If 

the dot is double, the location of the 13 components of its 

bisecting contour is included in the form of a 213 vector 

of double precision floating point numbers. From this point 

on, double dots are treated as two individual single dots. 

Then, for each single dot, the information contained in Table 

I is provided. Another of the outputs of the dot analysis tool 

is an image that illustrates all the information enumerated 

above in a graphic manner superimposed on the original 

Figure 17. Illustration of the operation of the analysis tool: a original scanned single dot, b result of 

analysis of single dot, c original scanned double dot, and d result of analysis of double dot. The type of 

black frame surrounding the dot corresponds to the type of dot. The dotted lines are the bisecting and 

tail-segmenting contours. The dashed lines are the fitted ellipses. The white crosses are the main dot and 

tail/satellite centroids. 



Table II. Parameters for the five different print modes. 

Print mode No. Directionality Carriage speed 

1 Unidirectional 15 ips 

2 Bidirectional 15 ips 




image. Figure 17 shows sample input and output images for 

both single and double dots. The output images show the 

ellipse fitted to the dot, the tail-segmenting contour, and the 

centers of mass of the main dot and of the tail. 

In order to allow for controlled variation of the dot 

characteristics, the test targets had to be encoded into Printer 

Control Language (PCL) commands. PCL commands embed 

printing attributes such as print resolution, carriage 

speed, and print directionality into the print job before 

sending it to the printer. The process of encoding a page in 

the PCL language consists of breaking the image file into 

horizontal stripes with height equal to the height of a swath. 

Then, each image file is converted to a PCL file that specifies 

the carriage speed, directionality of the print, resolution, and 

the number of nozzles to use. The PCL files corresponding 

to each of the image swaths are then sent sequentially to the 

printer by means of a proprietary software tool that allows 

the horizontal offset between swaths to be changed in steps 

as small as 1/13 of 1/600 in. 

Two printing attributes were varied throughout to obtain 

different dot characteristics: print speed and print directionality. 

A total of five different print modes were created. 

The parameters of each of the print modes are listed in Table 

II. A specific class of dots corresponds to each of these print 

modes. In order to identify the main differences between the 

type of dot produced by each print mode, the test target was 

printed and subsequently analyzed with the dot analysis tool. 

Effect of Print Speed on Dot Characteristics 

The first source of variability tested was the variability from 

pen to pen. Using the dot analysis tool, we were able to 

establish that the attributes of the printed dot remain more 

or less constant for a given print speed throughout a fairly 

large population of pens. We tested a population of 30 different 

pens and measured the characteristics of the printed 

dots for the 60 ips, bidirectional print mode. Figure 18 

shows the resulting fraction of dots with a tail (measured as 

number of tails divided by number of dots) and dot aspect 

ratio (measured as the ratio of the ellipse’s major to minor 

axes) for the pen population. Upon inspection of the plots, it 

becomes clear that there is not a significant variation of the 

dot characteristics from pen to pen, for a particular print 

mode. 

Figure 19 shows the average dot profile for the right-toleft 

swaths at different print speeds. It becomes evident from 

the inspection of these images that as carriage speed increases, 

the average dot elongation increases and satellites 

and tails tend to grow. Figure 20 shows the effect of speed on 

the average dot aspect ratio and the fraction of dots with a 

tail and corroborates quantitatively the qualitative assertions 

concluded from the inspection of Fig. 19: as print speed 

increases, the average dot aspect ratio increases and the fraction 

of dots with a tail increases. 

PSYCHOPHYSICAL EXPERIMENTS ON ALIGNMENT 

PERCEPTION 

Psychophysical experiments allow us to draw conclusions 

about perception. The objective of this section is to make 

inferences about the effect of dot characteristics on perceived 

alignment from responses of human subjects in constant 

stimuli and signal detection experiments. The five print 

modes described in the section “Output of Dot Analysis 

Figure 18. Statistics for sample pen population averaged across all dots in the test pattern for each pen: a 

fraction of dots with a tail and b average dot aspect ratio. 



Figure 19. Average dot profiles for different print speeds: a 15 ips, b 30 ips, c 45 ips, and d 60 ips. 

As carriage speed increases, average dot elongation increases, and the likelihood of tails and satellites 

increases. 

Tool” were used to print the test images shown to the subjects. 

The following sections describe the design and the 

results of the experiments. 

Constant Stimuli Test 

Recall from the section “Misalignment Measurement” that 

misalignment is measured as the average offset of the horizontal 

centroids in each column of dots in the upper swath 

with respect to the horizontal centroids in each column of 

dots in the lower swath, while taking into account the effects 

of skew (see Fig. 10). In this experiment, printed misalignment 

values ranging from 0/600 in. to 1.6/600 in. are chosen. 

Preliminary tests showed that this range was informative 

enough for our purposes since it contains values that are 

consistently perceived as aligned, consistently perceived as 

misaligned, and values that do not offer a clear choice. Thus, 

offset values that produce measured misalignment ranging 

from 0/600 in. to 1.6/600 in. were chosen to be tested. The 

actual measured misalignment values tested vary from print 

mode to print mode, since the only parameter we can 

change is the relative offset between swaths. A test image is 

printed for each of the offset values and shown to the subject. 

For this experiment, two test pages consisting of linebased 

drawings were used as test images (see Fig. 21). In 

order to measure printed misalignment for each test image, 

five test patterns arranged horizontally across the whole 

width of the page (see Fig. 22) were placed directly below 

each of the images and printed on the same page. The test 

patterns were hidden prior to the execution of the experiment. 

Image misalignment was estimated by averaging the 

misalignment across the five patches, and only images with 

alignment whose standard deviation across the five test 

patches was smaller than 0.1/600 in. were kept. The order of 

presentation was randomized and the subject was asked to 

answer whether he/she was able to detect misalignment in 

each of the test pages. A total of 16 subjects with normal or 

corrected to normal vision, who were students and/or staff 



Figure 20. Effect of print speed on a average dot aspect ratio and b 

fraction of dots with a tail. As print speed increases, the average dot 

aspect ratio increases and the fraction of dots with a tail increases. 

Figure 21. Test images used in constant stimuli test: a 600 dpi resolution 

image and b 300 dpi resolution image. 

members at Purdue University, participated in this 

experiment. 

The first test image was the 600 dpi resolution flowchart 

depicted in Fig. 21(a). Eleven versions of this image were 

printed with the 15 ips, unidirectional print mode, each version 

at a different misalignment value, for a total of 11 images. 

The second test image was the 300 dpi resolution flowchart 

depicted in Fig. 21(b). Eleven versions of this image 

were printed with each of the four remaining print modes, 

each version at a different misalignment value, for a total of 

44 images. Therefore, the total number of stimuli for the 

experiment was 55. Each subject was free to change the 

viewing distance to the page and to take as much time as 

needed to give a response. However, it was found that the 

subjects tended to hold the pages at a viewing distance of 

10 to 12 in., and that the average time to complete the experiment 

was less than 30 min. 

The proportion of “Detected” responses across subjects 

for each misalignment amount was recorded and plotted 

against the corresponding misalignment value. The data 

Figure 22. Arrangement of test patterns used to measure misalignment on test pages. These test patterns were 

printed below the images shown in Fig. 21 and were hidden during the psychophysical experiments. Figure 7 

shows the detailed structure within each of the test patterns. 



Figure 24. Average proportion of “Detected” responses across 16 subjects 

for a 45 ips bidirectional and b 60 ips bidirectional print modes. 

Standard probit analysis cannot be applied since the data points are not 

monotonic. 

Figure 23. Average proportion of “Detected” responses across 16 subjects 

and corresponding psychometric curves for a 15 ips unidirectional, 

b 15 ips bidirectional, and c 30 ips bidirectional print modes. 

Both estimated parameters and and the corresponding standard 

estimation errors are included. 

points were fitted with a cumulative Gaussian distribution 

by estimating the mean and standard deviation via Probit 

Analysis. 31,32 In this case, is related to sensitivity to 

changes in alignment: the larger its value, the less sensitive 

the subjects are. The parameter reflects both sensitivity to 

changes in alignment and response bias. Specifically, higher 

sensitivity leads to smaller values of . At the same time, 

however, the value of maydependonthesubject’sresponse 

criterion. For example, if the subject is conservative, 

that is, if he/she decides to the answer “Not Detected” when 

in doubt, will be larger. 

Figure 23 shows the resulting curves and data points 

from the experiments corresponding to three print modes: 

15 ips unidirectional, 15 ips bidirectional, and 30 ips bidirectional. 

Note that, as expected, the proportion of “Detected” 

responses increases as the misalignment value increases. 

This suggests that the point of perceived perfect 

alignment (the point at which the proportion of “Detected” 

responses is close to zero) coincides with the point of measured 

perfect alignment (the point at which measured misalignment 

is 0 in.). The plots include the estimated values 

for and as well as the standard error for each of the 



Figure 25. Average proportion of “Shifted to the Right” responses across 

ten subjects and corresponding psychometric responses for symmetric test 

with a 45 ips bidirectional and b 60 ips bidirectional print modes. 

Both estimated parameters and and the corresponding standard 

estimation errors are included. 

parameters. Note that cannot be estimated reliably because 

there are almost no data points for which the proportion of 

detections is near 0.5. Most of the data points correspond to 

a proportion of “Detected” responses equal to 0 or 1. Therefore, 

only can be used as a measure of sensitivity, although 

it may confound sensitivity with response bias. From the 

graphs, we can conclude that the higher print speed leads to 

lower sensitivity. 

Figure 24 shows the resulting data points from the experiments 

corresponding to the two remaining print modes: 

45 ips bidirectional and 60 ips bidirectional. Note that the 

proportion of “Detected” responses was close to zero for 

measured misalignment that was not 0 in.: between 0.4/600 

and 0.9/600 in. for the 45 ips print mode, and at 

1.5/600 in. for the 60 ips print mode. This suggests that the 

point of perceived perfect alignment does not correspond to 

the point of measured perfect alignment. This is related to 

Figure 26. New psychometric curves for a 45 ips bidirectional and b 

60 ips bidirectional print modes. Both estimated parameters and and 

the corresponding standard estimation errors are included. 

the fact that at higher print speeds, the dots are highly elongated 

and the dot’s centroid does not correspond to the 

perceived center of the dot. Since the data points do not 

exhibit the monotonicity characteristic of a Gaussian curve, 

Probit Analysis cannot be applied directly. 

In order to estimate the point at which alignment is 

perceived as perfect, a new set of constant stimuli tests was 

designed. For this experiment, vertical lines composed of 

two line segments with measured offsets near the points at 

which the psychometric curves reach their minimum value 

(0.75/600 in. for 45 ips and 1.50/600 in. for 60 ips) were 

printed. Seven values were chosen for the 45 ips print mode 

and ten values were chosen for the 60 ips print mode. A test 

pattern like the one in Fig. 7 was placed directly below the 

vertical line and printed on the same page to enable misalignment 

measurement. The test pattern was hidden prior 

to the execution of the experiment. The order of the presentations 

was randomized and the subject was asked to answer 

whether the lower segment was shifted to the left or to the 



Table III. Stimulus response matrix. 

Yes 

No 

Large misalignment Hits Misses 

Small misalignment False alarms Correct rejections 

Table IV. Signal detection test results. 

Table V. Estimated DL from signal detection test results. 

S 

1 / 600 in. 

mean DL 

1 / 600 in. 

stddev DL 

1 / 600 in. 

15 ips bidirectional 0.24 0.17 0.04 




mean d stddev d mean c stddev c 

15 ips bidirectional 1.45 0.27 −0.04 0.14 



60 ips bidirectional 1.57 0.25 0.00 0.25 

right with respect to the upper segment. A total of ten subjects 

with normal or corrected to normal vision, who were 

students and/or staff members at Purdue University, participated 

in this experiment. Once again, the subjects were allowed 

to change the viewing distance to the page and to take 

as much time as needed to give a response. Subjects took on 

average less than 15 min to complete the test. 

The proportion of “Shifted to the Right” responses 

across subjects for each misalignment amount was recorded 

and plotted against the corresponding misalignment value. 

The data points were fitted with a cumulative Gaussian distribution 

by estimating the mean and standard deviation 

via Probit Analysis. The mean value of the fitted Gaussian 

curves in this symmetric design is the point of subjective 

equality (PSE), that is, the point of measured alignment at 

which the line is subjectively perceived to be aligned over a 

large number of trials. The PSE provides a better estimator 

of the point of perfect perceived alignment than the misalignment 

value at which the propotion of “Detected” responses 

is minimum in the plots depicted on Fig. 24. Figure 

25 shows the resulting psychometric curves, along with their 

respective estimated parameters. The plots include the estimated 

values for and as well as the standard error for 

each of the parameters. These results demonstrate that the 

point of perceived perfect alignment does not correspond to 

the point of measured perfect alignment for the two print 

modes under consideration, as expected from the previous 

experiment. 

Now that we have a good estimator for the PSE, we can 

go back to the results in Fig. 24 and study them properly. 

The PSE might be thought of as the new origin for the data 

points of the constant stimuli tests for the 45 and 60 ips 

print modes depicted in Fig. 24: as we move away from the 

PSE (0.73/600 in. for the 45 ips print mode and 

1.64/600 in. for the 60 ips print mode) in either direction, 

the proportion of “Detected” responses increases. Thus, relocating 

the origin of the plots in Fig. 25 to the position of 

the PSE and plotting the data points at their absolute distance 

from the PSE results in a monotonic sequence, which 

allows the application of Probit Analysis. This is consistent 

with the 15 ips and 30 ips cases, in which the PSE is near 

0 in., and the data points exhibit a monotonic behavior as 

we move away from the origin. Figure 26 depicts the new 

psychometric curves for the original tests for 45 and 60 ips 

bidirectional, with the origin shifted to the position of the 

PSE and the data points located at their absolute distance 

from the PSE. Note that the value of is considerably higher 

for the 60 ips case than for any other case (see Fig. 23 and 

Fig. 26). This suggests that subjects might be less sensitive to 

changes in alignment at this particular print speed. 

Signal Detection Test 

The Gaussian parameter estimators from the constant 

stimuli test might be affected by noise for a variety of reasons, 

including response bias (the tendency of a subject to 

respond “Detected” or “Not detected” for reasons other than 

the percept of the stimulus itself) and lack of informative 

data points (those for which the proportion of “Detected” 

responses differs from 0 and 1). The latter is a consequence 

of the finite resolution of the printing device, which only 

allows us to change alignment in fixed-size steps. Signal detection 

tests are an alternative to measure a subject’s sensitivity 

(the equivalent to in the constant stimuli tests) that 

is less affected by response bias. 33 

The signal detection experiment we performed falls in 

the class of Yes-No experiments for sensitivity measurement. 

In particular, we are interested in measuring the ability to 

distinguish between two misalignment values, rather than 

the ability to detect the presence of misalignment, as in the 

constant stimuli test. To this end, test pages consisting of 

vertical lines were encoded in the PCL language. A test pattern 

like the one in Fig. 7 was placed directly below the 

vertical line and printed on the same page to enable misalignment 

measurement. The test pattern was hidden prior 

to the execution of the experiment. Two groups of test pages, 

each consisting of 20 pages, were printed with each of the 

print modes. The test images in one of the groups had 

smaller misalignment values than those in the other group. 

The standard deviation of the misalignment values within 

each group was less than 10% of the difference between the 

average alignment values of the two groups. The order of the 



presentations was randomized and the subject was asked 

whether the page belonged to the large misalignment group 

or not, one page at a time. A total of seven subjects with 

normal or corrected to normal vision, who were students 

and/or staff members at Purdue University, participated in 

this experiment. Each of the subject’s responses was tabulated 

into a stimulus response matrix (see Table III). The 

subjects were free to choose the most appropriate viewing 

distance to the test pages, and to take as long as they desired 

to evaluate each page. Subjects took on average less than 

20 min to go through the 40 images. 

Since there are a total of 20 images in each group, the 

number of hits plus the number of misses equals 20 as well 

as the number of false alarms plus the number of correct 

rejections. Therefore, it is only necessary to work with two of 

the four numbers in order to obtain all pertinent information 

about a subject’s performance. The following is a short 

description of the data analysis procedure. 33 

The hit rate H is the proportion of large misalignment 

trials to which the subject responded “Yes,” and the false 

alarm rate F is the proportion of small misalignment trials 

to which the subject responded “Yes.” A common measure 

of sensitivity in signal detection theory is d. It is defined in 

terms of the inverse of the normal distribution function, z, 

as 

d = zH − zF. 

The sensitivity measure d is unaffected by response 

bias. This is because if the subject has the inclination to give 

a particular answer, both zH and zF move in the same 

direction, e.g., if the subject gives preference to the “Yes” 

response, both zH and zF increase, but their difference 

does not change. The subject’s preference to a particular 

response, or the response bias c, is estimated as follows: 

8 

Figure 27. Interswath junctures of line segments consistently perceived as 

aligned for a 45 ips bidirectional swaths displaced by 0.7/600 in. 

and b 60 ips bidirectional swaths displaced by 1.5/600 in. print 

modes. The dashed lines correspond to the horizontal positions at which 

the vertically projected absorptance profiles in Fig. 28 take on the values 

0.9 for the innermost red lines, 0.65 for the middle green lines, and 

0.3 for the outermost blue lines. Scanned at 8000 dpi with QEA 

System. 

Figure 28. Absorptance profiles of interswath junctures that were consistently 

perceived as aligned for a 45 ips bidirectional and b 60 ips 

bidirectional print modes. Three selected absorptance levels are highlighted 

with dotted lines corresponding to the identical colored lines in 

Fig. 27. Note that both edges of the two interswath junctures intersect at 

the absorptance level 0.65 indicated by the green dotted line for both 

print modes. 



c =− 1 zH + zF. 

2 9 

Table IV contains the d and c results averaged across 

seven subjects. The values of c close to zero mean that there 

was little influence of bias on the recorded responses. Larger 

values of d imply higher sensitivity to detect the particular 

difference in stimulus magnitude, here, the difference in 

alignment between the small misalignment and the large 

misalignment. Note that the differences in misalignment between 

the small and the large misalignment groups are different 

for each print mode. Therefore, in order to compare 

different print modes with respect to sensitivity, a parameter 

called difference threshold (DL) has to be calculated from 

d. 

The parameter DL is defined as the smallest difference 

in stimulus magnitude that can be reliably detected. If the 

Gaussian distribution is the correct model for the psychometric 

function, DL corresponds to the difference between 

stimuli magnitudes S that produces d=1.So,ifd is proportional 

to S, DL is computed as follows: 

DL = S 

d . 

10 

The DL values estimated for each of the four print 

speeds are listed in Table V. It can be seen that the sensitivity 

of subjects to detect differences in alignment for the 15, 30, 

and 45 ips print modes is about the same, but the sensitivity 

for the 60 ips print mode is substantially lower (DL is 

larger). 

Note that some knowledge of how subjects perceive 

alignment is required for proper design of the signal detection 

test. In particular, an appropriate value of S is necessary 

for the test results to be meaningful. This is because if 

S is too large, the subject may not produce any errors, and 

it would not be possible to estimate d. On the other hand, 

if S is too small, the subject would perform at a chance 

level and the estimator would yield d=0. The information 

of how large S should be chosen to be for each print mode 

is readily extracted from the constant stimuli test results. A 

good rule of thumb is to pick S between and 2, where 

corresponds to the standard deviation of the psychometric 

curve from the constant stimuli tests. 

Another important fact we had to keep in mind when 

designing the signal detection tests was that for the 45 and 

60 ips, the PSE was not 0 in. In order for the signal detection 

test results to be meaningful, the misalignment values of 

both of the groups of prints must have the same sign relative 

to the PSE (they must both be located either to the right or 

to the left of the PSE). Otherwise, the resulting d would be 

an underestimation of the subject’s sensitivity. In fact, in the 

extreme case, the estimator would yield d=0 even if the 

subject could reliably discriminate the two stimuli. For example, 

if misalignment values of 0.2/600 in. and 

1.1/600 in. were chosen for the case shown in Fig. 24(a) 

(which corresponds to the constant stimuli test results for 

the 45 ips print mode), the subjects would judge the two 

levels as equally misaligned in a signal detection test, even 

though they are perceived as different: one is misaligned to 

the left and the other one is misaligned to the right. 

Discussion 

Point of Perceived Perfect Alignment 

Some insight to the fact that the point of perceived alignment 

differs from that of measured perfect alignment can be 

gained by examining an actual interswath juncture that was 

consistently perceived as aligned by the subjects. As the 

alignment values change, the appearance of each separate 

swath remains unchanged, but the relative horizontal positions 

of adjacent swaths change. Therefore, subjects make 

their decision as to whether the print is aligned or not based 

on the appearance of the interswath junctures. This fact was 

corroborated by the subjects after each of the sessions. Figure 

27 shows sample scanned interswath junctures from the 

images that were consistently perceived as aligned by the 

subjects for both 45 and 60 ips print modes. Recall that even 

though the horizontal center of mass of the segment from 

the upper swath is different from that of the segment from 

the lower swath, these particular arrangements were consistently 

perceived as aligned. To better understand the reasons 

why this happens, it is helpful to examine the average normal 

profiles of the upper and lower swath segments. 

Figure 28 shows the vertically projected absorptance 

profiles for the line segments that belong to the right-to-left 

and left-to-right swaths in both of the junctures depicted in 

Fig. 27. In both cases, the profiles show the asymmetry that 

results from tails and satellites that trail the main dots on the 

side opposite the direction of movement of the pen. Note, 

however, that in spite of the asymmetry of the profiles, the 

points at which they intersect lie on a horizontal line at 

approximately the same magnitude of absorptance in both 

cases (see green dotted line in Fig. 28). This level of absorp- 

Figure 29. Measured alignment versus perceived alignment for 45 ips 

bidirectional print mode. Measured misalignment increases from left to 

right and from top to bottom. The white cross indicates the location of the 

centroid in each average dot profile. 



tance corresponds to a straight vertical line in the scanned 

interswath junctures, which has also been highlighted with 

the middle (green) dotted line in Fig. 27. This suggests that 

the main cue to perceived alignment is the position at which 

the edges of the lines reach a certain locally averaged level of 

absorptance and, more specifically, the absorptance level 

0.65 highlighted by the green dotted line. Notice that this 

absorptance level corresponds roughly to the 60% threshold 

of the transition from the paper to the line peak absorptance 

levels. This threshold has been reported to be the one that 

defines the line width perceived by the human observer. 34 

For reference, two other levels of absorptance have been 

highlighted as well. 

Figure 29 shows a simulation of dot-level relationships 

between swaths, specifically, at the interswath juncture. The 

average dot profile for a single direction at a print speed of 

45 ips was calculated. To account for the opposite directionality 

of the pen at the interswath junctures, the profile was 

flipped horizontally and the flipped version was placed right 

below the original profile. The two profiles are displaced 

with respect to one another to illustrate the effect of dot 

elongation on the relation between perceived and measured 

alignment. The amount of the relative displacement increases 

from left to right and from top to bottom in steps 

that correspond to the response that they elicited: from misalignment 

values that were consistently perceived as misaligned, 

to values that were occasionally perceived as misaligned, 

to values that were consistently perceived as aligned. 

The first image in the sequence shows the relationship 

between two dots with perfectly aligned horizontal centroids 

in a configuration that was consistently perceived as misaligned. 

The next image in the sequence shows the situation 

where the horizontal centroids are displaced 0.37/600 in. 

with respect to one another in a configuration that was occasionally 

perceived as aligned. The last image in the top row 

illustrates the situation where the horizontal displacement 

between the dot centroids equals 0.75/600 in. This is the 

offset that was consistently perceived as aligned by the subjects 

for this particular print mode. The bottom row illustrates 

displacements that continue to increase, starting from 

the offset that was reliably perceived as aligned and ending 

with an offset that was again reliably perceived as misaligned. 

Figure 30 shows a similar sequence in coarser steps 

for the 60 ips print mode. These sequences of images are an 

alternative way of visualizing the fact that the main cue to 

perception of alignment is not the offset between centroids, 

since zero offset between dot centroids does not guarantee 

that the dot configuration will be perceived as aligned. 

Rather, subjects appear to base their decision on the overall 

dot shape including tails or satellites. 

Sensitivity to Changes in Alignment 

The constant stimuli test results allowed us to estimate two 

important parameters of alignment detection: the point of 

perceived perfect alignment and the sensitivity to detect differences 

in alignment. The estimation of the sensitivity via 

constant stimuli tests is not reliable for the reasons explained 

Figure 30. Measured alignment versus perceived alignment for 60 ips 

bidirectional print mode. Measured misalignment increases from left to 

right. The white cross indicates the location of the centroid in each average 

dot profile. 

earlier. This raised the need for signal detection tests that 

provide a means to reliably measure sensitivity. The results 

showed that subjects are less sensitive to changes in alignment 

with the 60 ips print mode than with any other mode. 

CONCLUSIONS 

We presented a combination of automated image analysis 

methods and psychophysical tests to shed light on the issue 

of how swath-to-swath ink jet alignment is perceived by the 

average observer. We developed algorithms to measure misalignment 

as printed on a page and to classify printed dots 

based on their characteristics. Using the tools we developed, 

we showed that dot variability from pen to pen is negligible. 

We demonstrated that the way alignment is perceived is 

highly dependent on the characteristics of the individual 

dots. As print speed increases, dot elongation increases and 

the presence of artifacts like tails and satellites becomes more 

evident. At small print speeds, dot shape tends to be symmetric 

about its centroid, and alignment of dot centroids 

corresponds roughly to alignment of dot outlines. At higher 

print speeds, dot shape becomes asymmetric about the dot 

centroid. In these cases, perfect alignment is not achieved by 

aligning dot centroids, but rather by aligning outlines at a 

certain level of absorptance. For the printer manufacturer, 

this implies that there is a need to develop alignment techniques 

that are based on alignment of ink outlines rather 

than on alignment of absorptance centroids. This conclusion 

corresponds to the results reported by Ward et al., 35 where 

the authors concluded that the subjects primarily used virtual 

edges to judge misalignment between two random dot 

clusters. 

Recall that the just noticeable angular offset between 

two line segments is called Vernier acuity and that it ranges 

from 5 to 10 seconds of arc. The sensitivity thresholds for 

perception of changes in alignment reported in this paper 

are within the order of the Vernier acuity: 0.2/600 in., the 

estimated DL for the 15, 30 and 45 ips print mode with the 

signal detection test, corresponds to 5.7 seconds ofarcata 

viewing distance of 12 in. On the other hand, 0.4/600 in., 

the estimated DL for the 60 ips print mode with the signal 

detection test, corresponds to 11.5 seconds of arc at a view- 



ing distance of 12 in. It is important to emphasize, however, 

that these thresholds do not remain unchanged as printing 

speed changes. Specifically, the sensitivity threshold is noticeably 

higher when carriage speeds go beyond 45 ips. This 

result was corroborated by the results of the psychophysical 

tests and corresponds to results reported by Patel et al., 18 

where the authors found that Vernier thresholds increase for 

dots with irregular shapes. 

ACKNOWLEDGMENTS 

The authors wish to thank Stuart Scofield, Bret Taylor, and 

Steve Walker of HP Vancouver, WA for their invaluable assistance 

and encouragement during the performance of this 

research. This work was supported by the Hewlett-Packard 

Company. 

REFERENCES 

1 “Document B 123: NP 13660. Measurement of image quality attributes 

for hardcopy output”, 7th Working Draft, ISO/IEC, Geneva, Switzerland, 

1995. 

2 Y. Kipman, “Image quality metrics for printers/plotters”, Proc. IS&T/ 

SID’s Fourth Color Imaging Conference (IS&T, Springfield, VA, 1995) 

pp. 134–138. 

3 Y. Kipman, “Image quality metrics for printers and media”, in Proc. 

IS&T’s PICS: Image Processing, Image Quality, Image Capture, Systems 

Conference (IS&T, Springfield, VA, 1998) pp. 183–187. 

4 D. Wolin, K. Johnson, and Y. Kipman, “The importance of objective 

analysis in image quality evaluation”, Proc. IS&T’s NIP 14 (IS&T, 

Springfield, VA, 1998) pp. 603–606. 

5 D. Aultman and R. Dumas, “Automating jet quality analysis using a 

scanner-based system”, Proc. IS&T’s NIP 20 (IS&T, Springfield, VA, 

2004) pp. 378–382. 

6 D. Forrest, J. Briggs, and M. Tse, “Print quality analysis as a QC tool for 

manufacturing inkjet print heads”, Proc. IS&T’s NIP 14 (IS&T, 


7 D. Wolin, K. Johnson, and Y. Kipman, “Automating image quality 

analysis”, Proc. IS&T’s NIP 14 (IS&T, Springfield, VA, 1998) pp. 

627–630. 

8 J. Briggs, M. Murphy, and Y. Pan, “Banding characterization for inkjet 

printing”, Proc. IS&T’s PICS: Image Processing, Image Quality, Image 

Capture, Systems Conference (IS&T, Springfield, VA, 2000) pp. 84–88. 

9 Y. Kipman and K. Johnson, “Dot placement analysis using a line scan 

camera and rigid body rotation”, Proc. IS&T’s PICS: Image Processing, 

Image Quality, Image Capture, Systems Conference (IS&T, Springfield, 

VA, 2001) pp. 68–71. 

10 P. Fleming, J. Cawthorne, F. Mehta, S. Halwawala, and M. K. Joyce, 

“Interpretation of dot area and dot shape of inkjet dots based on image 

analysis”, J. Imaging Sci. Technol. 47, 394–399 (2003). 

11 S. Wang, “Aerodynamic effect on inkjet main drop and satellite dot 

placement”, Proc. IS&T’s NIP 14 (IS&T, Springfield, VA, 1998) pp. 5–8. 

12 J. P. Allebach, “DBS: Retrospective and future directions”, Proc. SPIE 

4300, 358–376 (2001). 

13 D. Kacker, T. Camis, and J. Allebach, “Electrophotographic process 

embedded in direct binary search”, IEEE Trans. Image Process. 11, 

243–257 (2002). 

14 F. Baqai and J. Allebach, “Halftoning via direct binary search using 

analytical and stochastic printer models”, IEEE Trans. Image Process. 12, 

1–15 (2003). 

15 J. Lee and J. P. Allebach, “Inkjet printer model based halftoning”, IEEE 

Trans. Image Process. 4, 674–689 (2005). 

16 J. Yen, M. Carlsson, M. Chang, J. Garcia, and H. Nguyen, “Constraint 

solving for inkjet print mask design”, J. Imaging Sci. Technol. 44, 

391–397 (2000). 

17 D. Regan, Human Perception of Objects, 1st ed. (Sinauer, Toronto, 2000) 

pp. 45–49. 

18 S. Patel, H. Bedell, and M. Ukwade, “Vernier judgments in the absence 

of regular shape information”, Vision Res. 39, 2349–2360 (1999). 

19 W. Jang and J. P. Allebach, “Characterization of printer MTF”, J. 

Imaging Sci. Technol. 50, 264–275 (2006). 

20 J. Grice and J. P. Allebach, “The print quality toolkit”, J. Imaging Sci. 

Technol. 43, 187–199 (1999). 

21 N. Otsu, “A threshold selection method from gray-level histograms”, 

IEEE Trans. Syst. Man Cybern. 9, 62–66 (1979). 

22 A. Bjork, Numerical Methods for Least Squares Problems (Society for 

Industrial and Applied Math, Philadelphia, 1996). 

23 K. Fukunaga, Introduction to Statistical Pattern Recognition, 2nded. 

(Academic, Boston, 1990). 

24 M. Turk and A. Pentland, “Eigenfaces for face recognition”, J. Cogn 

Neurosci. 3, 71–86 (1991). 

25 V. Caselles, F. Catte, T. Coll, and F. Dibos, “A geometric model for active 

contours”, Numer. Math. 66, 1–31 (1993). 

26 M. Kass, A. Witkin, and D. Terzopoulos, “Snakes: Active contour 

models”, Int. J. Comput. Vis. 1(4), 321–331 (1987). 

27 V. Leavers, Shape Detection in Computer Vision Using the Hough 

Transform, 1st ed. (Springer, New York, 1992). 

28 H. Yuen, J. Illingworth, and J. Kittler, “Detecting partially occluded 

ellipses using the Hough transform”, Image Vis. Comput. 7, 31–37 

(1989). 

29 A. Fitzgibbon, M. Pilu, and R. Fisher, “Direct least square fitting of 

ellipses”, IEEE Trans. Pattern Anal. Mach. Intell. 21, 476–480 (1999). 

30 R. Halif and J. Flusser, “Numerically stable direct least squares fitting of 

ellipses”, Proc. Sixth Int’l Conf. Computer Graphics and Visualization, 

Vol. 1, 125–132 (1998). 

31 G. Gescheider, Psychophysics: Method, Theory, and Application, 2nded. 

(Erlbaum, Hillsdale, NJ, 1985). 

32 D. Finney, Probit Analysis, 3rd ed. (Cambridge University Press, New 

York, 1971). 

33 N. Macmillan and C. Creelman, Eds., Detection Theory: A User’s Guide, 

2nd ed. (LEA, New York, 2005) pp. 3–50. 

34 P. Engeldrum, “Quality of dot formed images”, Proc. SPIE 310, 118–124 

(1981). 

35 R. Ward, C. Casco, and R. Watt, “The location of noisy visual stimuli”, 

Can. J. Psychol. 39, 387–399 (1985). 




Characterization of Red-Green and Blue-Yellow 

Opponent Channels 

Bong-Sun Lee † 

School of Electrical and Computer Engineering, Purdue University, West Lafayette, IN 47907 

E-mail: bongsun.lee@thomson.net 

Zygmunt Pizlo 

Department of Psychological Sciences, Purdue University, West Lafayette, IN 47907 

JanP.Allebach 

School of Electrical and Computer Engineering, Purdue University, West Lafayette, IN 47907 

Abstract. The responses of opponent channels have been modeled 

in the past as a linear transformation of cone absorption values 

L, M, S. The authors asked two related questions: (i) which form of 

transformation is psychologically most plausible and (ii) is a linear 

transformation the right model, in the first place. The authors tested 

positions of unique hues for seven subjects in an xy chromaticity 

diagram as well as in a Boynton–MacLeod chromaticity diagram in 

log-coordinates. The results show that neither of the two opponent 

channels can be adequately approximated by a single straight line. 

The red-green channel can be approximated by two straight lines. 

The blue-yellow channel can be approximated by a quadratic function, 

whose middle section coincides closely with the daylight locus. 

These results show that linear models do not provide an adequate 

description of opponent channels. Our further analysis shows that 

there is a correlation between the red and the green directions. 



INTRODUCTION 

A trichromatic theory has been dominant in the field of 

color vision since the time of its formulation. It was originally 

proposed by Thomas Young 1 and then popularized by 

Helmholtz. 2 According to this theory, there are three receptors 

in the human eye that produce color sensations of blue, 

green, and red. Other colors are produced by combinations 

of these three. Despite its success in accounting for various 

color phenomena, the theory has failed to explain some important 

phenomena such as color blindness, simultaneous 

color contrast, color afterimages, etc. These color phenomena 

are explained by the opponent process theory proposed 

by Hering. 3 According to the opponent process theory, color 

is coded in the visual system in three channels: red-green, 

blue-yellow, and bright-dark. Green is a negative red, and 

blue is a negative yellow. As a result, no color appears simultaneously 

both red and green or blue and yellow. The theory 

received considerable attention after it had been tested and 

† 

The author is currently working at Thomson Inc., Burbank, CA. 

Received May 30, 2005; accepted for publication Aug. 26, 2006. 

1062-3701/2007/511/23/11/$20.00. 

confirmed by Hurvich and Jameson’s binocular fusion 

experiment. 4 

There are currently two theories accounting for the opponent 

color mechanisms. One postulates three stages and 

the other postulates two (Hurvich and Jameson suggested 

that two stages are sufficient 5 ). Vision science and the psychophysics 

community use a three-stage theory: (1) LMS 

cone excitation, (2) cone-antagonistic processing that can be 

derived as a linear transformation of the first stage, and (3) 

a higher-order chromatic mechanism of the coneantagonistic 

information. For the third stage, there exist two 

different chromatic mechanisms to obtain unique red and 

unique green, and one single mechanism for unique blue 

and unique yellow. 6–8 A more detailed description of the 

three-stage theory can be found in the recent work by 

Wuerger, Atkinson, and Cropper. 8 The two-stage theory is 

widely used in imaging systems research and 

applications. 9–18 Essentially, this theory takes the first two 

stages from the three-stage theory and ignores the third one, 

assuming that the third stage contributes little. By doing this, 

the two-stage theory is computationally quite simple because 

it does not include the nonlinear transformation of the third 

stage. But the computational simplicity of the two-stage 

theory comes at the price of providing a less accurate description 

of the color coding in the human visual system. A 

natural question is whether the approximation errors in the 

two-stage theory are justifiable. For example, if the errors are 

smaller than individual variability, then eliminating these errors 

will have no practical consequences for the color imaging 

industry. The main motivation behind our study is to 

provide empirical results that shed light on this question. 

Our results show that an accurate description of each 

observer’s color judgments requires the third stage, as suggested 

by the three-stage theory. Furthermore, we provide 

evidence showing that the blue-yellow channel cannot be 

modeled by a single straight line (contrary to Wuerger et al.’s 

finding). Finally, our results strongly suggest that these nonlinearities 

are not small as compared to individual variability 

and, therefore, should be included in imaging applications, 

23

Lee, Pizlo, and Allebach: Characterization of red-green and blue-yellow opponent channels 

such as image quality prediction, compression, broadcasting, 

color management, etc. 

The human visual system acquires spectral information 

by means of three types of cones with maximum sensitivity 

in the, long, medium, and short wavelengths L,M,S. This 

information is then represented by the responses of three 

opponent channels. There has been a substantial amount of 

research on the transformation between the responses of 

cones (LMS) and the responses of opponent channels 

(OPP). Most transformations have been assumed to be linear 

and are represented as a 33 matrix. There is also another 

device-independent space that is used to represent colors. 

This is the CIE XYZ space. 19 CIE XYZ is often used in 

engineering applications. The transformations between XYZ 

and OPP and between XYZ and LMS have also assumed to 

be linear by Smith and Pokorny, 20 Stockman, MacLeod, and 

Johnson, 21 and Stockman and Sharpe. 22 

The nature of color representation is critical in the color 

imaging industry. When images are processed, one tries to 

minimize the perceived error between the transformed and 

the original images. It is, therefore, important to know 

which space provides an adequate representation of the color 

percept, so that the error can be computed in this space. 

Visible difference predictors are computational tools that try 

to accomplish just that. 16–18 Color images are often represented 

in the CIE XYZ space. But the percept uses the OPP 

space. Therefore, the visible difference predictors have to 

transform the former into the latter before the perceived 

difference is computed. As pointed out just above, the imaging 

community assumes that a linear transformation can 

be used to characterize the relation between these two 

spaces. However, if a linear transformation is not adequate, 

the visible difference will not be predicted accurately. Another 

application of OPP space is found in the compression 

of images to utilize the fact that the human eye is not as 

sensitive to chromatic values as it is to luminance. Therefore, 

the human visual system can afford to lose more information 

in the chrominance signals than in the luminance signal. 

In this application, precise decomposition of images 

into opponent colors is a key factor for more effective and 

optimal compression results. 

There has been growing evidence indicating that the 

linearity assumption may not be valid. 23–29 Before we discuss 

details of the violations of the linear model, we introduce 

five different transformations between XYZ and OPP and 

highlight differences and similarities among them. These 

transformations will be called Zhang, 9 Hurvich, 10 Flohr, 11 

Hunt, 12–14 and Wandell. 14,15 Each of these transformations is 

represented as follows: 

The columns of A provide isoluminant and isochrominant 

modulations in the CIE xy chromaticity diagram: C l 

specifies the direction of isochrominant modulation and C rg 

and C by define the directions of isoluminant modulation. 

For the isoluminant modulation that isolates the red-green 

opponent mechanism, the response of the blue-yellow 

mechanism is supposed to be zero. Similarly, the red-green 

mechanism response to the blue-yellow stimulus also should 

be zero. The two opponent-color directions specified by the 

vectors C rg and C by for each of the five transformations are 

shown in Fig. 1. The vector C rg is the direction of the redgreen 

channel and C by is the direction of the blue-yellow 

channel. Unique spectral hues (green, blue, yellow) as identified 

by Hurvich and Jameson 4 for subject DJ are represented 

in Fig. 1 as circles. 

The transformations of Hurvich and Flohr are almost 

identical. Let us assume that Hurvich and Jameson’s unique 

hues shown in Fig. 1 are an adequate representation of colors 

in the human visual system. (Recall, however, that there 

is individual variability with respect to unique hues.) Then, 

the transformation by Hurvich and Flohr seems to be the 

best for both blue and yellow. However, in the case of green, 

the model by Hunt or Wandell would be better than the one 

by Hurvich and Flohr. Clearly, none of these transformations 

seem to be adequate for all three unique hues. 

The question of psychophysical plausibility of linear 

models has been examined by Larimer, Krantz, and 

Cicerone 6,7 Burns, Elsner, Pokorny, and Smith, 23 Ayama, Nakatsue, 

and Kaiser, 24 Ikeda and Uehira, 25 Chichilnisky and 

Wandell, 26 Zaidi, 27 Webster, Miyahara, Malkoc, and 

Raker, 28,29 and Wuerger et al. 8 Some measurements were 

performed with a mixture of monochromatic lights and others 

with computer generated stimuli. 

Larimer et al. 6 reported that the blue-yellow opponent 

channel (red-green equilibria) satisfied Grassmann-type additivity 

laws. Specifically, the combination of unique blue 

X 

1 

1 

Y AO O 2 C l C rg C by O O 2 

Z= O 3= O 3, 

1 

where C l , C rg , and C by are the columns of A. The matrices 

are described in detail in Appendix A.1 (available as Supplemental 

Material on the IS&T website, www.imaging.org). 

Figure 1. Opponent-channel directions of five opponent-channel matrices 

in the CIE xy chromaticity diagram. The point “W” is the equal energy 

“white.” 



and unique yellow remains an equilibrium color (neither red 

nor green). However, another experiment conducted by 

them showed nonlinear additivity in the red-green opponent 

system (blue-yellow equilibria). 7 A similar result was described 

in the most recent study by Wuerger et al. 8 From the 

findings of nonlinearity of unique red and unique green in 

cone space, they postulated that there are three chromatic 

mechanisms required to account for the four unique hues: 

two color mechanisms that yield unique red and unique 

green, respectively, and one chromatic mechanism for 

unique blue and unique yellow. Burns et al. noted from two 

observers that constant hue loci were typically curved (this is 

called the Abney effect 30 ) in the chromaticity diagram. 23 

Each of their unique hue loci was fairly straight except the 

curved unique blue locus. However, their unique reds were 

not collinear with unique greens. Similar results are found in 

Valberg’s determination of four unique hue curves. 31 

Rather than using the mixture of monochromatic lights 

as in Larimer et al., Burns et al., Ikeda and Uehira, and 

Ayama et al.’s experiments, Chichilnisky and Wandell used 

stimuli generated on a computer monitor. 26 They also concluded 

that the opponent classification was not linear and 

described it by using a piecewise linear model. Recently, 

computer generated stimuli have been widely used to measure 

not only the loci of unique hues, but also the loci of 

constant hues, 32,33 and all the loci look similar to those of 

previous findings with monochromatic color stimuli. 34,35 

A recent study by Webster et al. is quite representative 

for the current understanding of the relation between the 

cone absorptions and opponent channels. 28,29 They measured 

the direction of unique hues for several subjects. In 

their experiment, the initial adaptation to the gray background 

lasted 3 min, and the intertrial adaptation lasted 3s. 

The color stimulus was presented for 280 ms. Each trial presented 

moderately saturated stimuli. Figure 2 shows the 

opponent-color directions of one of their subjects in the xy 

chromaticity diagram. The red-green direction cannot be 

approximated by a single straight line. The same is true for 

the blue-yellow direction. 

In this paper, we provide a further test of the linearity 

assumption for the transformation from LMS to OPP and 

XYZ to OPP. Similar studies on the unique hue characterization 

have been done in the past and the results described 

in this paper conform to those former findings. Compared 

to previous studies, our psychophysical experiment used 

more subjects, and the exposure duration was unlimited. 

Using unlimited exposure duration more closely approximates 

natural viewing conditions. Our discussion focuses on 

the correlation between the red and the green directions, and 

on the relation between the daylight locus and the entire 

blue-yellow channel. It is important to point out that when 

werefertoopponentcolors,werefertocolorsintheopponent 

(perceptual) color space, rather than to colors in the 

chromaticity diagram. Specifically, in the chromaticity diagram, 

opponent colors do not have to lie on a single straight 

line going through the point representing an achromatic 

color. In fact, they do not. 

Figure 2. Opponent-color directions of one subject in the xy chromaticity 

diagram measured by Webster et al. p. 1548, Fig. 2, observer EM. 29 

Before testing the subjects in the main experiment, we 

had them perform the standard color deficiency tests. 

COLOR DEFICIENCY TEST 

Subjects 

We tested five male observers (SL, WJ, OA, GF, KL), one of 

whom is the first author of this paper, and two female observers 

(BZ, YB). We used two tests: Ishihara’s test for color 

deficiency 36 and the Farnsworth–Munsell 100-hue test. 37 

Five observers wore normal untinted glasses for their eyesight 

correction. Both tests were done under daylight D65 

simulated by a viewing booth (GretagMacbeth SpectraLight 

II, 617 Little Britain Road, New Windsor, NY 12553). Subjects 

SL, BZ, WJ, GF, and KL had perfect scores for all 25 

Ishihara plates, while YB and OA responded incorrectly on 

some plates. Specifically, OA responded incorrectly on plate 

#19, and YB on plates #5, 7, 9, 12, 16, 17, and 22. According 

to the instruction for the interpretation of the test result, 

YB’s color vision is not regarded as normal, but since she 

read 15 plates out of the first 21 plates normally, she cannot 

be treated as a color deficient, either. In fact, her result of the 

Farnsworth-Munsell 100-hue test indicated she had good 

color discrimination ability as described next. With a more 

sophisticated color vision test such as anomaloscope, YB was 

found to have a normal color vision. 

The Farnsworth–Munsell 100-hue test directly measures 

the subject’s ability to perform general color discrimination. 

It enables subjects with normal color vision to be categorized 

into the classes of superior (total error score is less than 

16), average (total error score from 16 to 100), and low (total 

error score greater than 100) color discrimination. Figure 3 

shows the results for each subject. Subjects SL, GF, and KL 

achieved a zero error score on this test. These subjects also 

scored perfectly on the Ishihara test. BZ and WJ’s color vision 

can also be treated as superior since they had only one 

wrong arrangement of purplish colors and cyanish colors, 

respectively (their error score was 4). Recall that these 

subjects obtained perfect scores on the Ishihara test. Subjects 

YB and OA gave more incorrect answers with error score 

16, and thus can be categorized as normal observers with 



Figure 3. Results of the seven subjects in the Farnsworth-Munsell 100-hue test. Perfect performance is represented 

by a perfect circle. The number of bumps represents the number of errors, and the height of the bumps 

represents the magnitude of each error. For example, OA made four errors. 

average color discrimination ability compared to the other 

subjects. 

EXPERIMENT 

General Methods 

Apparatus and Stimuli 

A calibrated cathode ray tube (CRT) computer monitor 

(EIZO FlexScan T965, EIZO Nanao Technologies Inc., 5710 

Warland Drive, Cypress, CA 90630) was used to display 

stimuli in a darkened room. The calibration was done by a 

PR705 spectroradiometer (Photo Research Inc., 9731 Topanga 

Canyon Place, Chatsworth, CA 91311-4135) using a 

procedure similar to that described by Berns, Motta, and 

Gorzynski. 38 To evaluate the performance of the calibration, 

125 patches using the combination of five digital values (0, 

32, 96, 145, 245) were generated and tested. CIE color differences 

were computed between the measurements from the 

patches and predictions from the calibration. The results 

showed average E * ab =0.5 and maximum E * ab =1.62. 

The subject viewed the stimuli from a distance of approximately 

20 in. A square patch with size 22 deg 2 was 

shown at the center of the monitor. The background was 

uniform neutral gray color with the chromaticity value of 

(x=0.33, y=0.33) and a luminance value of 20 cd/m 2 . 

Ten stimuli were generated for each color (red, green, 

blue, yellow) by using different mixtures of the red, green, 

and blue phosphors of the CRT monitor. For example, to 

generate the ten red stimuli that were used to find unique 

red, each patch was generated by setting the red phosphor to 

its maximum intensity, the green phosphor to one of ten 

evenly spaced digital values ranging from 15% to 85% of its 

maximum, and the blue phosphor to a random value. These 

ten values for the green phosphor produced ten different 

levels of saturation of the red patch. This mixture looked 

red, but not necessarily unique red. The range of the random 

* 

setting of the blue phosphor corresponded to ±15E ab units 

around the unique red of subject SL (this range contained 

unique reds of the remaining subjects). The subject’s task 

was to adjust the intensity of the blue phosphor to make the 

mixture look unique red (i.e., neither bluish red nor yellowish 

red). Unique green was determined the same way as 

unique red. In the case of unique blue, the subject adjusted 

the intensity of the red phosphor to cancel green for ten 

different saturations of blue. Unique yellow was determined 

similarly by asking the subject to adjust the intensity of the 

green phosphor to cancel red. Note that the luminance of 

stimuli with different saturations and hues was not the same. 

It is known that luminance has little or no effect on the color 

settings chosen as unique. 6,7 We directly tested this assumption 

in a control experiment described in Appendix A. 2 



(available as Supplemental Material on the IS&T website, 

www.imaging.org). 

Figure 4. Subject SL’s settings of unique red medium saturation using 

four different durations of adaptation, a variation in chromaticity x and 

b variation in chromaticity y, as function of trial number. 

Procedure 

There were four sessions. Each session consisted of ten trials 

and tested only a single color, i.e., red, green, blue, or yellow. 

The individual trials measured unique hue at different levels 

of saturation. In each trial, the subject viewed the patch in 

the center of the CRT monitor. A slide-bar was provided on 

the monitor for the subject to adjust the color of the patch. 

Changing the position of the slide-bar changed the intensity 

of one phosphor (the other two phosphors stayed at the 

initial setting). At the beginning of each trial, the phosphor 

intensity corresponding to the center position of the slidebar 

was randomized to prevent the subject from using this 

position as cue to color. The subject’s task was to adjust the 

patch to a unique hue. Each trial was preceded by 3 min of 

adaptation to a neutral gray background. This duration was 

chosen based on the results of a preliminary experiment, 

which is described next. Each trial lasted about 1 to 2 min. 

One session lasted about 1h. Each subject was limited to 

one session a day. 

Note that since each trial lasted up to 2 min, the subject’s 

visual system adapted to the color displayed. As a result, 

this color was not constant throughout the trial, but was 

nevertheless close to the unique hue that the subject was 

supposed to produce. For example, in a trial where unique 

red was produced, the patch had only a small component of 

blue or yellow. Therefore, it is reasonable to expect that the 

adaptation changed the appearance of the patch with respect 

to the red component, but not much with respect to the blue 

or yellow components. The effect of adaptation (if present) 

is likely to lead to increased variability of judgments from 

trial to trial, but not to systematic errors because (i) the 

initial intensity of the variable phosphor was random and 

(ii) the ten levels of saturation were presented in random 

order. This conjecture was verified in a control experiment 

described in the next section. 

Preliminary Experiment 

The first author was the subject. The subject repeated ten 

trials for the same stimulus but with different durations of 

adaptation between trials: 0, 1, 3, and 5 min. In each trial, SL 

viewed the stimulus (medium saturated red) after he 

adapted to the neutral gray background for the given period 

of adaptation. As in the main experiment, the subject’s task 

was to adjust the patch to a unique hue by changing the 

position of the slide-bar. 

Figure 4 shows the subject SL’s settings of unique red 

using four different durations of adaptation. It is seen that 

the 0 and 1 min adaptation periods produce systematic 

changes in the perceived color of the stimulus: the values of 

x and y systematically increase with the trial number. Specifically, 

the slope of the regression line is significantly different 

from zero p0.05. On the other hand, 3 and 5 min 

of adaptation produce no systematic changes in the perceived 

color. The slope of the regression line was not significantly 

different from zero. Therefore, in the main experiment 

we used a 3 min adaptation between trials. To further 

minimize the effect of adaptation, the stimuli were presented 

in a random order in the main experiment. 

Results 

Figure 5 shows the opponent-channel directions for each 

subject in the xy chromaticity diagram. It can be seen that 

the red-green channel cannot be represented by a single 

straight line. The red and the green parts of this channel can 

be approximated by straight line segments, but these segments 

meet at an angle different from 180°. Similarly, the 

blue-yellow channel cannot be represented by a straight line; 

a single curved line seems more appropriate. 

To verify whether a piecewise linear regression is adequate, 

we performed a regression analysis separately for 

each subject and each color. Specifically, linear and quadratic 

functions were used and the significance of the quadratic 

term was tested. The results for each subject are shown in 

Table I. 

It is seen that the quadratic term was significant for the 

blue part of the blue-yellow channel for six out of the seven 

subjects. An additional analysis shows that the quadratic 

term is significant for the blue-yellow channel in all subjects. 



Figure 5. Unique hues for each subject in xy chromaticity diagram. Redstars-greendiamonds and 

bluecircles-yellowcrosses. Individual data points represent settings in individual trials. “W” is the point that 

corresponds to equal energy white. The triangle in the diagram represents the gamut of the computer monitor 

that was used in the experiment. 

Therefore, we independently approximated the red and 

green settings by two straight line segments, and the blueyellow 

settings by a single quadratic function. Using two 

separate functions (a quadratic one for the blue channel and 

a linear one for the yellow channel) is likely to improve the 

fit, but at the expense of using more parameters. 

Figure 6 shows the results from Fig. 5, but with best 

fitting lines for blue-yellow (dash-dotted line), red-green 

(dashed line), and daylight locus (solid line) superimposed. 

The dashed lines for the red and green parts are the best 

fitting lines for the data. By the best fitting line, we mean a 

line minimizing the sum of squared distances of data points 

from this line in the direction orthogonal to the line. The 

actual computations were completed by singular value decomposition 

(SVD). In the conventional regression, the sum 

of squared differences in the direction of the variable to be 



Figure 6. Opponent-channel colors for each subject with the best fitting lines for blue-yellow dash-dotted line, 

red-green dashed line, and daylight locus solid line superimposed. 

predicted is minimized. However, in our analysis we are not 

interested in making predictions about either y or x. Therefore, 

there is no reason to minimize errors in either of these 

two directions. In fact, since the task was to adjust the hue so 

that it is unique, it is reasonable to assume that the errors 

can be adequately modeled by distribution along a direction 

orthogonal to the opponent-channel direction. Thus, the 

best approach seems to be the one that minimizes the sum 

of squares of shortest distances between the data points and 

the resulting straight line. This is done by determining the 

eigenvector that is associated with the larger eigenvalue for a 

given data set. The direction of this vector represents the 

slope of the best fitting line. The intercept is obtained by 

assuming that the line goes through the center of the gravity 

of the data set. 

For the blue-yellow data, we cannot do the same analysis 

because for quadratic regression, there is no direction 

that can be used as a common orthogonal direction for all 

data points. Instead, we performed the regression in a new 

coordinate system. We first compute the eigenvector for the 

blue-yellow data whose direction maximizes the variance. 

Then, we rotate the x-coordinate system in such a way that 



Table I. p-values for testing the significance of the quadratic term in the regressing 

line approximating the data for individual subjects. p-values less than 0.05 indicate 

that the quadratic term is statistically significant. 

R G B Y B-Y 

SL 0.92 0.05 0.05 0.05 0.05 

BZ 0.06 0.64 0.05 0.14 0.05 

WJ 0.25 0.50 0.05 0.06 0.05 

YB 0.44 0.06 0.05 0.10 0.05 

GF 0.96 0.05 0.05 0.08 0.05 

KL 0.24 0.06 0.25 0.13 0.05 

OA 0.55 0.83 0.05 0.97 0.05 

the new x axis coincides with this direction (we call the new 

coordinate system, x−y). As a result, y is nearly orthogonal 

to blue-yellow everywhere. Next, we run a quadratic regression 

that minimizes errors along the y direction. These 

lines are shown in Fig. 6. The daylight locus is a quadratic 

function computed by Judd. 39 It is drawn within the daylight 

phase: 4000–25 000 K. Note that for all subjects, the best 

fitting line for blue-yellow closely coincides with the daylight 

locus. 

Figure 7 shows the relationship between the angle R of 

the red part and the angle G of the green part of the redgreen 

opponent channel for each subject. Each angle was 

obtained by measuring the angle between the x axis and the 

fitted line in Fig. 6. The error bars in each direction indicate 

one standard deviation of the estimated angle. The orientation 

of the best fitting line in each graph coincides with the 

direction of the eigenvector that is associated with the larger 

eigenvalue for the given data. It is seen that there is a systematic 

relation between the two angles. The squared correlation 

coefficient is r 2 =0.38 [Fig. 7(a)]. If the data point 

representing subject OA is excluded (this data point is characterized 

by large standard errors), the squared correlation 

coefficient is substantially higher r 2 =0.83 [Fig. 7(b)]. 

Figure 8 shows the opponent-channel colors in a 

Boynton–MacLeod chromaticity diagram 40 in logcoordinates: 

logS/L+M versus logL/L+M space. 

The quantum absorption rates L, M, S were computed from 

the tristimulus values X, Y, Z using the matrix from Kaiser 

and Boynton. 19 The daylight locus (solid line) and best fitting 

lines for each opponent channel are superimposed. All 

lines in the graph were obtained via the same method used 

to determine the lines in the xy chromaticity diagram as 

shown in Fig. 6. In this space, the daylight locus seems to 

correspond to a straight line: The squared correlation coefficient 

computed from 20 equally spaced points from the 

daylight locus curve is equal to 0.99. Lee also observed that 

the daylight locus can be approximated by a straight line in 

log-log coordinates, 41 and this feature was used by Wei, 

Figure 7. Angle between the red line and the green line the unit is 

degrees: a for seven subjects G =1.14 R +142.0, r 2 =0.38 and b 

for the six subjects remaining after subject OA is excluded G =1.38 R 

+147.8, r 2 =0.83. Solid lines represent best fitting lines obtained by a 

regression on the given data. The error bars in each direction indicate ± 

one standard deviation of the estimated angle. 

Pizlo, Wu, and Allebach in their model for color constancy. 42 

Again, neither the blue-yellow nor the red-green direction 

can be adequately approximated by a single straight line. 

DISCUSSION 

For linear models of color vision to be adequate, it is necessary 

that the two equilibrium lines be straight lines in the xy 

chromaticity diagram. We (and others) have shown that 

these lines are not straight lines in xy space. It follows that 

linear models do not provide an adequate description of the 

opponent channels. 

As shown in Fig. 5, there is not much variability across 

subjects with respect to unique blue and unique yellow. On 

the other hand, there is some variability across subjects with 

respect to the directions of the lines representing unique red 

and unique green, although the two directions appear to be 

correlated. Figure 9 shows the distributions of maximally 

saturated unique hues measured by the seven subjects. Variations 

in unique blue and unique yellow are relatively small 

compared to those in unique red and unique green. The 



Figure 8. Opponent-channel colors for each subject in logS/L+M vs. logL/L+M space with blueyellow 

dash-dotted line, red-green dashed line, and daylight locus solid line superimposed. 

* 

range of unique hues in E ab units is 18 for red, 26 for 

green, 11 for blue, and 9 for yellow. 

In Fig. 6, we showed that the blue-yellow opponent hue 

locus closely coincides with the daylight locus. To identify 

the relation between the daylight locus and the entire blueyellow 

opponent channel, we rendered 20 blue and yellow 

Munsell chips under 10 daylight illuminants chosen from 

the range 4000–25 000 K. These 20 chips were selected from 

the Munsell book to adequately represent the unique blueyellow 

locus of the subject SL. Figure 10 shows the rendering 

results of 20 Munsell chips. The 20 circles represent 20 Munsell 

chips, a thick dotted line shows the daylight locus, and 

the several thin solid lines are the rendering results. It can be 

seen that the chromaticities of the light reflected by blue and 

yellow surfaces closely coincide with the loci of unique blue 

and unique yellow hues, as shown in our measurements (see 

Fig. 6) and in measurements presented in Gouras’ Plate 

10(b). 43 This similarity suggests that the blue-yellow channel 



Figure 9. Distributions of maximally saturated unique hues measured by 

seven subjects. Variations in unique blue and unique yellow circles are 

relatively small compared to those in unique red and unique green 

diamonds. 

Figure 10. Rendering of 20 blue and yellow Munsell chips under ten 

daylight illuminants chosen from the range 4000–25 000 K. The 20 

circles represent 20 Munsell chips, a thick dotted line shows the daylight 

locus, and the several thin solid lines are the rendering results. 

evolved to efficiently serve in solving color constancy problem: 

unique blue and unique yellow surfaces will look 

unique when rendered under any daylight. More generally, 

changing the daylight when a natural surface is viewed, 

changes the response of the blue-yellow channel, only. The 

red-green channel is invariant to such changes. 

Consider now the relation between our results and 

those of Webster et al. 28,29 Figure 8 is log scaled version of 

the Boynton–MacLeod chromaticity diagram. In the plot, 

even though the phases of daylight are represented by a 

straight line, the entire blue-yellow channel cannot be represented 

by a straight line. The same is true for the red-green 

channel. This characteristic of the unique hue directions 

confirms that these directions are not only off the theoretical 

cardinal opponent axes (the x and y axes of the Boynton– 

MacLeod diagram), but they also cannot be characterized by 

straight lines. In many respects, our results are similar to 

those of Webster et al. 28,29 . Their unique hue settings also 

showed a certain amount of departure from the cardinal axes 

(modified Boynton–MacLeod diagram). For none of their 

observers could the red-green channel be approximated by a 

single straight line; for some subjects, the same was true for 

the blue-yellow channel. This fact clearly shows that any 

simple linear transformation between OPP responses and 

LMS responses cannot describe the postprocessing stage of 

the human visual system. Webster et al. described the locus 

of unique hues as straight lines. Recall, however, that in our 

results, the blue part of the blue-yellow channel could not be 

approximated by a straight line. This difference between 

Webster et al.’s results and ours could be a consequence of 

the fact that we used a wider range of saturations. 

In our experiment, subjects viewed the stimulus for 

about one minute on the average. This amount of time was 

necessary for the subject to be able to adjust the color of the 

patch. A question remains as to the effect of this prolonged 

adaptation on the directions of the opponent channels. In 

other words, while the subject is adjusting the color to 

unique red, does the adaptation affect the saturation of red 

only, or the hue as well? To verify whether unique hues 

adjusted in the main experiment still look unique when the 

exposure duration of the color is short, subject SL ran four 

sessions, ten trials per session, one session for each hue. 

After initial adaptation to the gray background for 3 min, a 

sequence of ten trials began. In each trial, the stimulus was 

shown for 280 ms. There was a 3sadaptation period between 

trials. The trials involved the colors that subject SL 

chose as unique in the main experiment. The order of saturations 

was randomized. SL reported that the colors in each 

session still looked unique. Therefore, we conclude that the 

long exposure duration in our main experiment did not 

substantially affect the directions representing opponent 

channels. 

Finally, we would like to point out that the complex 

relation between cone responses and opponent channels 

suggests that it may be unwarranted to talk about color coding 

on the retina (as Young-Helmholtz theory claims). Color 

as a perceptual attribute is associated with the functioning of 

cortical areas of the brain, rather than of the retina itself. It 

is more appropriate to talk about coding spectral information 

on the retina. In fact, Hurvich and Jameson made this 

point very clearly. 44 

CONCLUSION 

In summary, our best fitting lines of unique hues in the xy 

chromaticity diagram, as well as in the log scaled version of 

the Boynton–MacLeod chromaticity diagram, reveal that a 

linear-model representation of opponent colors is not a precise 

characterization of the relation between OPP responses 

and LMS responses (or tristimulus XYZ). A simple nonlinear 

transformation (e.g., differentiable), however, does not 

seem to be plausible, either. Before such a model is proposed, 

one has to design psychophysical experiments that 

will allow reliable measurement of the relation between cone 

absorptions and responses of opponent channels for an arbitrary 

stimulus, not only for unique hues as was done in 

our study, as well as in prior studies. 



REFERENCES 

1 T. Young, On the Theory of Light and Colours (The Society, London, 

1802). 

2 H. von. Helmholtz, Treatise on Physiological Optics, translated from the 

3rd German ed. (Dover Publications, New York, NY, 1867/1925). 

3 E. Hering, Outlines of a Theory of the Light Sense, L.M.HurvichandD. 

Jameson, Trans. (Harvard University Press, Cambridge, MA, 1878/1964). 

4 L. M. Hurvich and D. Jameson, “The binocular fusion of yellow in 

relation to color theories”, Science 114, 199–202 (1951). 

5 L. M. Hurvich, Color Vision (Sinauer, Toronto, 1981). 

6 J. Larimer, D. H. Krantz, and C. M. Cicerone, “Opponent-process 

additivity-I: red/green equilibria”, Vision Res. 14, 1127–1140 (1974). 

7 J. Larimer, D. H. Krantz, and C. M. Cicerone, “Opponent process 

additivity-II. yellow/blue equilibria and nonlinear models”, Vision Res. 

15, 723–731 (1975). 

8 S. M. Wuerger, P. Atkinson, and S. Cropper, “The cone inputs to the 

unique-hue mechanisms”, Vision Res. 45, 3210–3223 (2005). 

9 X. Zhang and B. A. Wandell, “A spatial extension of CIELAB for digital 

color image reproduction”, J. Soc. Inf. Disp. 5, 61–63 (1997). 

10 L. M. Hurvich and D. Jameson, “Some quantitative aspects of an 

opponent-colors theory. II. brightness, saturation, and hue in normal 

and dichromatic vision”, J. Opt. Soc. Am. 45, 602–616 (1955). 

11 T. J. Flohr, B. W. Kolpatzik, R. Balasubramanian, D. A. Carrara, C. A. 

Bouman, and J. P. Allebach, “Model based color image quantization”, 

Proc. SPIE 1913, 270–281 (1993). 

12 R. W. G. Hunt, The Reproduction of Colour (Fountain, London, 1995). 

13 S. N. Pattanaik, J. A. Ferwerda, M. D. Fairchild, and D. P. Greenberg, 

“Multiscale model of adaptation and spatial vision for realistic image 

display”, Proc. of ACM, SIGGRAPH 1998, pp. 287–298. 

14 M. D. Fairchild, Color Appearance Models (Addison Wesley Longman, 

London, 1998). 

15 B. A. Wandell, Foundations of Vision (Sinauer, Toronto, 1995). 

16 S. Daly, “The visible differences predictor: an algorithm for the 

assessment of image fidelity”, Digital Images and Human Vision. A. B. 

Watson, Ed. (MIT Press, Cambridge, MA, 1993) pp. 179–206. 

17 J. Lubin, “A visual discrimination model for imaging system design and 

evaluation”, Vision Models for Target Detection and Recognition. E. Peli, 

Ed. (World Scientific, Singapore, 1995) pp. 245–283. 

18 W. Wu, Z. Pizlo, and J. P. Allebach, “Color image fidelity assessor”, Proc. 

IS&T PICS 2001 (IS&T, Springfield, VA, 2001) pp. 148–152. 

19 P. K. Kaiser and R. M. Boynton, Human Color Vision, 2nd ed. (Optical 

Society of America, Washington, DC, 1996). 

20 V. C. Smith and J. Pokorny, “Spectral sensitivity of the foveal cone 

photopigments between 400 and 500 nm”, Vision Res. 15, 161–171 

(1975). 

21 D. Stockman, I. A. MacLeod, and N. E. Johnson, “Spectral sensitivities of 

the human cones”, J. Opt. Soc. Am. A 10, 2491–2521 (1993). 

22 D. Stockman and L. T. Sharpe, “Spectral sensitivities of the middle- and 

long-wavelength sensitive cones derived from measurements in observers 

of known genotype”, Vision Res. 40, 1711–1737 (2000). 

23 S. A. Burns, A. E. Elsner, J. Pokorny, and V. C. Smith, “The Abney effect: 

Chromaticity coordinates of unique and other constant hues”, Vision 

Res. 24, 479–489 (1984). 

24 M. Ayama, T. Nakatsue, and P. K. Kaiser, “Constant hue loci of unique 

and binary balanced hues at 10, 100, and 1000 td”, J. Opt. Soc. Am. A 4, 

1136–1144 (1987). 

25 M. Ikeda and I. Uehira, “Unique hue loci and implications”, Color Res. 

Appl. 14, 318–324 (1989). 

26 E. J. Chichilnisky and B. A. Wandell, “Trichromatic opponent color 

classification”, Vision Res. 39, 3444–3458 (1999). 

27 Q. Zaidi, “Parallel and serial connections between human color 

mechanisms”, in Applications of Parallel Processing in Vision, J. R. 

Brannan, Ed. (Elsevier Science, Amsterdam, 1992). 

28 M. A. Webster, E. Miyahara, G. Malkoc, and V. E. Raker, “Variations in 

normal color vision. I. cone-opponent axes”, J. Opt. Soc. Am. A 17, 

1535–1544 (2000). 

29 M. A. Webster, E. Miyahara, G. Malkoc, and V. E. Raker, “Variations in 

normal color vision. II. unique hues”, J. Opt. Soc. Am. A 17, 1545–1555 

(2000). 

30 W. W. Abney, “On the change in hue of spectrum colours by dilution 

with white light”, Proc. R. Soc. London, Ser. A 83, 120–127 (1910). 

31 A. Valberg, “A method for the precise determination of achromatic 

colours including white”, Vision Res. 11, 157–160 (1971). 

32 P. C. Hung and R. S. Berns, “Determination of constant hue loci for a 

CRT gamut and their predictions using color appearance spaces”, Color 

Res. Appl. 20, 285–295 (1995). 

33 F. Ebner and M. D. Fairchild, “Finding constant hue surfaces in color 

space”, Proc. SPIE 3300, 107–117 (1998). 

34 D. L. MacAdam, “Loci of constant hue and brightness determined with 

various surrounding colors”, J. Opt. Soc. Am. 40, 589–595 (1950). 

35 A. R. Robertson, “A new determination of lines of constant hue”, AIC 

Proceedings Color 69 1, 395–402 (1970). 

36 S. Ishihara, Tests for Colour-Blindness (Kanehara Shuppen Co. Ltd., 

Tokyo, Japan, 1977). 

37 D. Farnsworth, The Farnsworth-Munsell 100-Hue Test for the 

Examination of Color Discrimination (Munsell Color, Macbeth Division 

of Kollmorgen Instruments Corp., New York, 1957). 

38 R. S. Berns, R. J. Motta, and M. E. Gorzynski, “CRT colorimetry. part I: 

theory and practice”, Color Res. Appl. 18, 299–314 (1993). 

39 G. Wyszecki and W. S. Stiles, Color Science: Concepts and Methods, 

Quantitative Data and Formulae, 2nd ed. (Wiley Interscience, New York, 

1982). 

40 D. I. A. MacLeod and R. M. Boynton, “Chromaticity diagram showing 

cone excitation by stimuli of equal luminance”, J. Opt. Soc. Am. 69, 

1183–1186 (1979). 

41 H.-C. Lee, “A computational model for opponent color encoding”, 

Advanced Printing of Conference Summaries, SPSE’s 43rd Annual 


42 G. Wei, Z. Pizlo, W. Wu, and J. P. Allebach, “Opponent channels and 

color constancy”, Invest. Ophthalmol. Visual Sci. 40, 749 (1999). 

43 P. Gouras, The Perception of Colour (Vision and Visual Dysfunction),Vol. 

6 (MacMillan, Oxford, 1991). 

44 L. M. Hurvich and D. Jameson, “Human color perception”, Am. Sci. 57, 

143–166 (1969). 




High Dynamic Range Image Compression by Fast 

Integrated Surround Retinex Model 

Lijie Wang, Takahiko Horiuchi and Hiroaki Kotera 

Graduate School of Science and Technology, Chiba University, Inage-ku, Chiba, Japan 

E-mail: lijiewang@graduate.chiba-u.jp 

Abstract. A novel compressing method of high dynamic range image 

based on fast integrated surround Retinex model is proposed in 

this paper. The proposed method has two novelties. First, multiscale 

surround images are integrated to a single surround field, which is 

applied to center/surround single-scale Retinex (SSR) model. The 

method reduces the “banding artifact” seen in normal SSR and simplifies 

the complicated computational steps in conventional multiscale 

Retinex. Second, the Gaussian pyramid method is introduced 

to cut the computation time for generating a large-scale surround by 

tracing a “reduction” and “expansion” sequences using down and up 

sampling followed by linear interpolation. The computational expense 

is dramatically saved less than 1/100 for getting a surround 

by Gaussian convolution with large kernel size. The proposed model 

worked well in compressing the dynamic range and improving the 

visibility in heavy shadow areas of natural color images while preserving 

pleasing contrast. 



INTRODUCTION 

Human vision is a complicated automatic self-adaptation 

system. It is capable of seeing over five orders in magnitude 

simultaneously and can gradually adapt to scenes with high 

dynamic ranges of over nine orders in magnitude. The current 

display devices, such as cathode ray tube (CRT), cannot 

capture the dynamic range more than 100:1. To recreate the 

viewer’s sensation of the original scene in current display 

devices, the high dynamic range (HDR) of the scene has to 

be compressed to the low dynamic range of the device. This 

is a difficult problem because the visual system is too complicated 

and current technique cannot yet understand it 

completely. 

The many published papers on HDR image compression 

are classified into two groups: Spatially-invariant tone 

reproduction curve (TRC) and spatially-variant tone reproduction 

operator (TRO) methods. 1 TRC operates pointwise 

on the image data which is actually based on the global 

adaptation of human vision. Algorithms by Tumblin et al., 2 

Tumblin and Rushmeier, 3 belong to this catagory. Pattanaik 

et al. 4 proposed a time-dependent method based on the time 

adaptation of human vision, which also uses the global adaptation 

models. TRO uses the spatial structure of the image 

data and attempts to preserve local image contrast. The algorithm 

by Chiu et al. 5 belongs to this catagory. TRC is 

Received Sep. 12, 2005; accepted for publication Jun. 6, 2006. 

1062-3701/2007/511/34/10/$20.00. 

simple and efficient, but at the expense of local contrast loss 

because of processing the whole image with a single curve. 

TRO, which is traditionally based on a multiresolution decomposition 

algorithm, such as Gaussian decomposition, 

works well in measuring and preserving local image contrast. 

However, any methods can model only a part of the 

complicated adaptation process of human vision. 

This paper follows the method of TRO, but presents a 

new idea based on the Retinex theory of the human vision 

process. Retinex is a typical method of TRO and has been 

broadly used in image processing, such as color image appearance 

improvement, 6,7 and also HDR image compression, 

e.g., by Carrato 8 who adopts a rational filter substituting 

for a Gaussian filter. Human vision can see the world 

without being affected by the spatially nonuniformity of illumination 

and the color of the illuminant, with what we 

call lightness and color constancies. Based on these characteristics, 

Land and McCann proposed Retinex. 9–14 Retinex is 

very useful in color image processing and has been improved 

during past forty years. Multiscale Retinex (MSR), generated 

by the weighted sum of multiple single-scale Retinex (SSR), 

is the most popular algorithm, because it can suppress the 

banding artifacts around high contrast edges in SSR. Since 

the optimization of weights is not easy, 15 conventional MSR 

simply applies equal weights to all scales of SSR but does not 

always give a satisfactory image. Kotera et al. 6,7 proposed an 

adaptive scale-gain MSR to improve the color appearance in 

conventional MSR, but the selection of scales and weights is 

still complicated, and the computation cost is too expensive. 

In this paper, a new fast and simple algorithm is proposed 

without banding artifacts caused by the conventional 

SSR model. The proposed algorithm adopts an integrated 

multiscale surround image composed of several luminance 

surround images to apply to the SSR model, which substitutes 

for the conventional integrated MSR composed of several 

SSR. The Gaussian pyramid is introduced to generate an 

integrated surround image quickly. The original image is 

repeatedly down sampled and divided by 2 in width and 

height, and the coarsest down-sampled image on the top of 

the pyramid is convoluted with the corresponding smallest 

size Gaussian filter, resulting in a surround image equivalent 

to the largest kernel size, so that the computational expense 

is dramatically reduced. By this model we get results comparable 

to the published papers in HDR image compression. 

In the following sections, first, we review the recent 

34

Wang, Horiuchi, and Kotera: High dynamic range image compression by fast integrated surround Retinex model 

progress in Retinex models. Next we propose the integrated 

surround Retinex algorithm, and third discuss the optimum 

parameters and improvement in speed. In addition, HDR 

image compression gives some examples which demonstrate 

good visibility in heavy shadow while preserving pleasing 

local contrast. Finally, we draw conclusions and insight into 

our future work. 

RETINEX MODEL 

The Retinex algorithm proposed by Land 9–13 is based on 

their Mondrian experiments and was improved by McCann 

et al. 16 It is a classical vision model with forty years history 

and recently received attention again. 17 Land suggested that 

color appearance is controlled by surface reflectance rather 

than by the distribution of reflected light and proposed three 

color mechanisms for the spectral responses of the cone 

photoreceptors. He called these mechanisms Retinexes because 

they are thought to be some combination of retinal 

and cortical mechanisms. 18 

According to Land, human visual system has the functions 

that recognize the world without being affected by spatially 

nonuniform distribution of illuminant. Basically Retinex 

is a model that eliminates the effect of the 

nonuniformity of illumination. Simply, the image I captured 

by camera is equivalent to the product of the reflectance R 

and illuminant distribution L. According to RI/L, we can 

restore reflectance R from Image I by inferring illumination 

L. 

Though various enhancements to the theory have been 

proposed, its key feature is that the Retinex algorithm explicitly 

treats the spatial distribution of illumination. According 

to the path-based model based on Mondrian experiments of 

Land and McCann, 10 the luminance difference of two separated 

points in the scene is obtained by the ratio of the 

neighboring points along the path. When gray step patches 

with linear reflectance are lit by the illumination which has 

the opposite gradient, the sequence of darkness appearance 

is not changed regardless of whether each patch reflects the 

same amount of light physically, if the relative luminance 

ratios on the boundaries of each edge are traced. To estimate 

the distribution of illumination L, various ways of taking 

paths into account have been published. The random walk 

model 18 computes the luminance product of each point 

from the distributed initial points in the image by a random 

walk. The Poisson model 19 approaches the spatial gradient in 

illumination from the change in the second derivative of the 

signal and computes it by inversion. McCann-Sobel model 20 

iteratively computes the luminance ratio along spiral paths 

while continuing to down-sample the image. Another iterative 

model by Funt 21 traces eight neighbors. The iterative 

model is a two-dimensional extension of the path-based 

model, where a new value is calculated for each pixel by 

iterative comparison. 

The center/surround model simply estimates the luminance 

L around a pixel in consideration by averaging the 

image I with Gaussian filter. Based on the work by Land, 13 

NASA (Refs. 22–26) developed MSR model by integrating 

multiple SSRs with different scales and weights. Furthermore, 

a quadratic programming method minimizes a second 

differential cost function by determining undefined Euler- 

Lagrange coefficients under the constraint of a spatial 

smoothing condition for image and illumination. Because 

the path-based model is complicated, the concise center/ 

surround model is selected in this paper. The reflectance 

image Rx,y is calculated by the ratio of center Ix,y to the 

surround Sx,y, simply noted as R=C/S. The spatial distribution 

of illumination Lx,y is equivalent to surround, 

which is calculated by averaging the original image Ix,y 

with a Gaussian filter. 

The most representative C/S MSR model of NASA is 

processed in logarithmic space. The following equations describe 

the process: 

i 

R MSR 

M 

x,y = 

m=1 

i 

w m R SSR x,y, m ; i = R,G,B, 

I 

i 

i x,y 

R SSR x,y, m = log 

I i x,y G m x,y ; i = R,G,B, 

G m x,y = K m exp− x 2 + y 2 / m 2 , 

G m x,ydxdy 

1 

2 

=1. 3 

Equation (2) expresses the output of SSR model as the ratio 

of the center pixel C=I i x,y to the surround S=I i G m , 

where G m denotes Gaussian averaging filter with scale m and 

standard deviation m and the symbol denotes convolution. 

The defect of SSR is a banding artifact appears around 

high contrast edges. A MSR model without banding artifact 

has been developed by Jobson et al., 22–26 integrating multiple 

SSRs with different standard deviations m and appropriate 

weight w m as expressed by Eq. (1). However, the optimization 

process of m and w m is unclear and these parameters 

must be decided by trial and error. In addition, logarithmic 

conversion accentuates the dark noise level in shadow region 

and the dynamic range expansion in the processed image 

needs to be limited. Furthermore, because the basic logarithmic 

model treats R, G, and B channels independently and 

the dynamic range of each channel is normalized to the 

range of the display device, the color balance cannot be 

maintained so that a wide uniform area in the image, such as 

sky or wall tends to a gray world. Jobson et al. 25 regulated 

the range of the output image by lower and upper clipping 

of the wide histogram. Rahman et al. 26 improved the color 

restoration with additional logarithmic terms corresponding 

to each color band signal divided by the sum of color band 

signals. They call this model multiscale Retinex with color 

restoration. Kotera et al., 6 proposed an adaptive scale-gain 

MSR model with stable and excellent color reproduction in 

linear space without using logarithmic conversion. In this 

model, the surround image generated only from the luminance 

image is used for the R, G, and B channels in common, 

which maintains the color balance. They also proposed 

an automatic setting method for weights adapted to the scale 



gain. However, since the computation for weights needs the 

histograms luminance SSRs corresponding to the multiple 

scales and takes too much time with increasing Gaussian 

kernel size, it still needs improvement for practical use. 

INTEGRATED-SURROUND RETINEX MODEL 

In this paper, we propose a concise new Retinex model different 

from the conventional MSR. Our work is mainly 

based on the work of Kotera et al. 6,7 First, we adopted linear 

space without logarithmic conversion to avoid instability for 

noise and output range spreading in dark shadows. Second, 

we used only the luminance channel to form the surround 

for each color channel in order to keep color balance. The 

major difference from Kotera’s method is that the new model 

creates an integrated multiscale luminance surround from 

multiple luminance surround images by Gaussian filters 

with different standard deviation m . The proposed model 

can suppress unwanted banding artifacts as well as the adaptive 

MSR model of Kotera. We introduced the Gaussian 

pyramid to produce the integrated surround image, by 

which the convolution computation for smoothing the original 

image with a Gaussian filter was dramatically reduced. 

The following subsection details improvements in our new 

algorithm. 

Integrated-Surround Retinex Algorithm 

Figure 1 illustrates the proposed integrated-surround Retinex 

model. Instead of the weighting sum of multiple SSRs, 

the proposed model integrates m=1M different surround 

images S m into a single surround image S sum with adaptive 

weight parameters w m . To keep color balance, S m is calculated 

by convoluting the luminance image Yx,y with the 

Gaussian filter G m with standard diviation m as Eq. (6) 

expressed. The output of Eq. (4) is the ratio of the center 

pixel I i to integrated luminance surround S sum and A is a 

gain coefficient which will be discussed detailed in the coming 

section on optimum parameters 

I i x,y 

SSR sum x,y, m = A 

S sum x,y, m ; 

i = R,G,B,A: gain coefficient, 

M 

S sum x,y, m = w m S m x,y, m , 

m=1 

4 

5 

where 

S m x,y, m = G m x,y Yx,y; 

m =2 m , Yx,y: luminance channel, 

M 

 

m=1 

w m =1. 

In the proposed method, M times of division is avoided 

in the computation of multiple SSRs and replaced with the 

easy summation instead. Figure 2(f) shows a sample obtained 

from the SSR process by the proposed method by 

integrating the three surround images of m =8,32,128 

with uniform weight of 1/3. It does not provide the dramatic 

improvement in shadow appearance as does NASA as 

shown in Fig. 2(d) or our previous adaptive scale-gain MSR 

in Fig. 2(e), but it suppresses the banding artifact very well 

in comparison with a conventional middle scale SSR in Fig. 

2(b) and is clearly better than the large scale SSR in Fig. 2(c). 

In addition, contrast appears more natural without over emphasis 

in comparison with NASA in Fig. 2(d) or our previous 

MSR in Fig. 2(e). 

Optimum Parameters 

The Retinex model aims to reproduce the original visual 

images, but in practice, the original scene is usually unknown 

unless the observer has seen the captured scene 

standing at the same place and the same time. Thus the 

setting of the optimum parameters is difficult without the 

original image. In this paper, as illustrated in Fig. 3, a test 

scene “color block” under nonuniform illumination in our 

laboratory is captured by a digital camera, then the camera 

image is modified using Adobe Photoshop by trial and 

error method until it is seen approximately matched to the 

visual scene. The modified image is taken as a target image. 7 

To make a quantitative estimation for the proposed 

model and find the optimum parameters, the color differences 

E ab between the visual target image and the pro- 

* 

cessed images are evaluated in CIELAB color space as 

follows: 

6 

7 

E * ab = L *2 + a *2 + b *2 1/2 , 8 

Figure 1. Proposed Retinex model using integrated surround. 



L * = L R * − L V * , a * = a R * − a V * , b * = b R * − b V * , 9 

where L * , a * , and b * are tristimulus values of CIELAB color 

space, R represents the results of proposed method, V represents 

target image 

L * = 116 f Y Y n −16, 

a * = 500 f X X n − f Y Y n, 

b * = 200 f Y Y n − f Z Z n, 

= ft t1/3 for t 10 

7.787t + 16/116 for t 0.008856, 

where X, Y, and Z are CIEXYZ tristimulus values and X n , Y n , 

and Z n are the CIEXYZ tristimulus values of the reference 

white point. Considering the computation expense and processing 

speed, it is hoped to produce a MSR image from a 

small number of SSRs. Empirically, to produce a MSR image 

without banding artifact, at least three SSR images are 

needed. As well, first, we used three scales M=3 of surround 

images, small 1 =2, middle 2 =16, and large 

3 =128 to get an integrated surround in the proposed 

method. Then we adjusted the weights w m to minimize 

the color difference between the target image C and the processed 

output for the camera image B in Fig. 3. Figure 4 

illustrates the results in the case of M=3. Because the possible 

number of combinations for the weights w m with 

gain parameter A becomes too large, we cut the unnecessary 

tests by observing the tendency of color difference changes 

corresponding to each combination. First fixing the weight 

w 1 to 0.1, with the condition w 1 +w 2 +w 3 =1,a 

combination of w 2 and w 3 is changed. Next fixing 

w 2 to 0.1, a combination of w 1 and w 3 is also 

changed. When the gain A=0.8 and the weights w 1 

=0.3, w 2 =0.1, and w 3 =0.6, the smallest color difference 

E * ab =8.6 is obtained. From the tendency of these color 

difference changes in Fig. 4, we can draw the conclusion that 

with the decrease in w 3 , the smallest color difference corresponding 

to each combination tends to increase and goes 

Figure 2. Sample by proposed Retinex model in comparison with conventional methods. 

Figure 3. Synthesis of target image visually matched to real scene. 

Figure 4. Color reproducibility by proposed model with three-scale sets 

m =2,16,128. 



up rapidly for w 3 0.5. Hence w 3 0.5 and large 

scale 3 =128 are necessary. We verified this condition again 

by fixing w 3 to 0.6 and 0.5, respectively, while changing a 

combination of w 1 and w 2 , and reached the same 

conclusion, which is almost the same as reported by 

Yoda et al. 7 

We can also draw another conclusion from the experiments, 

namely that w 1 is more important than w 2 in 

color reproduction, because the smallest color difference increased 

for w 2 w 1 when w 3 is fixed to around 

0.5. Thus we moved to the tests for the simpler case of two 

scales where the middle scale 2 =16 is discarded and a combination 

of small 1 =2 and large 3 =128 scales are 

used. The same test process is performed. Figure 5 illustrates 

the results in the case of M=2. When the gain: A=0.8, and 

weights: w 1 =0.4, w 3 =0.6, the best result E * ab =8.54 

is obtained, which is a little bit smaller than the case of three 

scales M=3, but considered to be almost the same color 

reproducibility as the result with three surround images. 

In addition, we also tested the color reproducibility for a 

different set of three scales ( 1 =8, 2 =32, 3 =128). As 

* 

illustrated in Fig. 6, the minimum color difference E ab is 

obtained when the gain A=0.8 and weights w 1 =0.2, 

w 2 =0.1, and w 3 =0.7, but it is a little bit worse than 

shown in Fig. 4 M=3 and Fig. 5 M=2. 

The typical resultant images are compared with NASA 

(d) and our previous adaptive scale-gain MSR (h) in Fig. 7. 

The best image with the smallest color difference for M=3 

by the proposed model is shown in Fig. 7(e) and that for 

M=2 in Fig. 7(f), respectively. In a tested color block image, 

banding artifacts are not seen in the reproduction by the 

proposed integrated-surround Retinex model using only two 

scales of luminance surround images. 

Improvement in Fast Computation 

The Retinex algorithm is very time-intensive due to a convolution 

between the original image and Gaussian filters in 

order to calculate surround images. Particularly, as the kernel 

size of the Gaussian filter increases, the computation 

time dramatically increases. The proposed model has the 

same problem, too. For example, when using a Gaussian 

filter with =128 (kernel size=4+1=513513 pixels) 

for the image size 1280960, it took more than one hour 

(Pentium 1 GHz, Memory 256 MB, MATLAB). For practical 

use, the time expense has to be reduced. Because time is 

mainly consumed in calculating the surround image, the 

Gaussian pyramid method is introduced to accelerate the 

Figure 5. Color reproducibility by proposed model with two-scale sets 

m =2,128. 

Figure 6. Color reproducibility by proposed model with three-scale sets 

m =8,32,128. 

Figure 7. Color reproducibility results by the proposed model in comparison with conventional methods. 



convolution speed in this paper. The Gaussian pyramid substitutes 

a large-scale convolution for a very small-scale one 

through up/down-sampling and interpolation sequences. 

Accordingly, the time expense is dramatically reduced. 

The convolution process in Gaussian pyramid is illustrated 

in Fig. 8. First, the original luminance image g 0 x,y is 

placed at the bottom, and each successive higher level is a 

smaller version scaled down by 1/2 in width and height of 

the previous level. Through the K step sequences, image 

group: g 1 ,g 2 ,...,g K is constructed. The image in level k is a 

copy reduced in resolution by 2 −k of the image g 0 x,y in 

level 0, which characterizes the multiresolution pyramid 

structure. The up process from g 0 to g 1 ,...,g K is finished by 

down-sampling the low-pass image by a Gaussian filter with 

half the rate. 

In this paper, we used a low pass filter with coefficients 

w=0.0500 0.2500 0.4000 0.2500 0.0500 approximated to 

Gaussian, which is circularly symmetric without half-pixel 

offsets. It works very rapidly because it is symmetric and 

applied separately in the horizontal and vertical directions. 2 

Designating the 1/2 reduction function as Reduce, we express 

the upward down-sampling Gaussian pyramid by Eq. 

(11), 

g k = Reduceg k−1 = Downsample 1/2 Lowpassg k−1 

Lowpassg k−1 = m g k−1 ; means convolution. 

m = m ij = w i · w j ; i,j = 1,2, ... ,5 

S K = g K G m x,y, K , 

S k−1 = Expands k = Upsample 2 Interpolates k ; 

k = K,K −1, ...,1. 

12 

13 

The surround S m expressed in Eq. (6) can be substituted by 

S 0 , and according to the Gaussian pyramid, S 0 can be obtained 

by the K-step up-sampling process after convoluting 

g K with the Gaussian filter G m K . Because the sizes of both 

g K and G m K are reduced to 2 −K 2 −K , the computation 

time is dramatically reduced. To avoid the loss of original 

image information, in this paper the minimum image size of 

the top level K image obtained by the down-sampling process 

is limited to 3232. 

Table I gives examples of the computation time before 

and after Gaussian pyramid for two different size images. 

For the original image g 0 with size of 256192, the size of 

top image g 2 is reduced to 6448 after K=2 steps down 

sampling. Because of m = K 2 K , in this case of K=2,we 

need to compute the convolutions for K =2,4,8,16,32, 

equivalent to m =8,16,32,64,128, respectively. For m 

=64 and 128, before and after Gaussian pyramid the computation 

time is reduced to about 1/10 and 1/15, respectively. 

The time is further reduced with increasing m .For 

larger image size, 1280960, after K=4 steps downsampling, 

the size of top image g 4 is reduced to 8060. As 

Table I(b) illustrates, we need only to compute K =2,4,8, 

equivalent to m =32,64,128, respectively. The computation 

w = w i = 0.05,0.25, . 0.4, . 0.25, . 0.05: 

lowpass filter coefficients. 

11 

When the reduced image g k at the required level K is obtained, 

convolution with a small-sized Gaussian filter with 

standard deviation K creates the reduced surround image 

S K corresponding to level K. Then S K is expanded to twice in 

width and height by interpolation and up sampled at twice 

the rate. The process is repeated until the surround image S 0 

with the same size as the original image is obtained. This 

downward up-sampling process is expressed by Eqs. (12) 

and (13), 

a 

Table I. Reduction in process time by Gaussian pyramid. 

Scale 

256192 

process time s 

Image size 

2561926448 


m m Normal Pyramid 

3 8 0.29 0.24 

4 16 0.75 0.24 

5 32 2.40 0.39 

6 64 9.13 0.90 

7 128 166.3 10.65 

Image size 

b 

Scale 

1280960 


12809608060 


m m Normal Pyramid 

5 32 59.10 5.13 

6 64 236.1 5.34 

Figure 8. Fast computation method for surround by Gaussian pyramid. 

7 128 4118 9.29 



time is reduced to about 1/10, 1/45, and 1/450 after the 

pyramid, respectively. The computation time is even more 

dramatically reduced not only with increasing m , but also 

with increasing image size. As shown in Table I(b), for image 

size 1280960, the computation time is reduced to 1/443 

for m =128 after pyramid. 

Since Gaussian pyramid processing uses the coarsest 

down-sampled image version of the original image for computation 

of the surround image, whether the Retinex image 

quality is affected or not has to be re-estimated. Again we 

evaluated the color difference between the resultant images 

after Gaussian pyramid and the target visual image color 

block. As shown in Fig. 9, in the case of M=3 with 

m =2,16,128, the smallest color difference E * ab =8.54 is 

obtained when gain A=0.65, w m =0.1,0.1,0.8. Aswell, 

for the case of M=2 with m =2,128 in Fig. 10, the smallest 

color difference E * ab =8.5 is obtained when gain A 

=0.6, w m =0.1,0.9. We also tested m =8,32,128 

equivalent to K =2,8,32 for the same condition as subsection 

Optimum Parameters. Figure 11 illustrates the results. 

We obtain almost the same color reproduction accuracies 

through Gaussian pyramid processing. 

Figure 12 gives some examples before and after Gaussian 

pyramid with the same parameters. The resultant image 

with Gaussian pyramid is much the same as the results without 

Gaussian pyramid. As visually observed in Fig. 12(a) 

through (f), the three pairs of resultant images for [A=0.5, 

w m =1/3, m =8,32,128], [A=0.6, w m =0.1,0.1,0.8, 

m =8,32,128], and [A=0.8, w m =0.2,0.1,0.7, 

m =8,32,128] resulted in much the same image appearance 

with and without Gaussian pyramid, and bear comparison 

with NASA in (h). Because the true target image is 

unknown in this outdoor scene, the optimal parameters may 

be different from those of test target image color block. The 

proposed system resulted in the excellent rendition (i) even 

for the default parameters, A=0.5, w 1 =w 2 =0.5, 

m =2,128 with Gaussian pyramid. 

HIGH DYNAMIC RANGE (HDR) IMAGE 

COMPRESSION 

The proposed model also worked well for HDR image compression. 

Considering the computation time, we again 

adopted the pyramid process to create the surround image. 

We did not need any particular postprocess for normal LDR 

images after Retinex process to regulate the dynamic range. 

But for the most HDR images, a postprocess is necessary for 

displaying them onto normal LDR display devices. Here the 

luminance channel is also applied to compute the surround 

for our HDR image compression in order to maintain color 

balance. First, we compute the integrated surround Retinex 

image Y R x,y for HDR luminance channel by 

Y R x,y = Yx,y 

S sum 

. 14 

Then we make use of Y R to obtain the condition for 

compressing the HDR image to LDR image for the display 

device. We found that the histogram of Y R is mostly concentrated 

in the lower range, while scattered in the middle to 

higher ranges for our tested HDR images as illustrated in 

Fig. 13. Thus we divided the higher range of Y R by large 

interval and the lower range by small interval not to lose the 

details. First, the histogram of Y R is divided into two parts 

[Min-Mean] and [Mean-Max] by the mean value Mean. 

Second, the pixel numbers Num 1 less than Mean and Num 2 

larger than Mean are calculated respectively. Third, the ratios 

of Num 1 and Num 2 to all pixel numbers are calculated by 

Eqs. (15) and (16). Then, the bins are calculated by Eq. (17), 

ratio 1 = 

Num 1 

Num 1 + Num 2 

, 

15 

ratio 2 = 

Num 2 

Num 1 + Num 2 

, 

16 

bin 1 = 255 ratio 1 ; bin 2 = 255 ratio 2 . 17 

Figure 9. Color reproducibility by proposed pyramid with three-scale 

sets m =2,16,128. 

Figure 10. Color reproducibility by proposed pyramid with two-scale 

sets m =2,128. 

Figure 11. Color reproducibility by proposed pyramid with three-scale 

sets m =8,32,128. 



Figure 12. Samples by the proposed model. 

Then the two ranges of [Min-Mean] and [Mean-Max] are 

uniformly divided into bin 1 and bin 2 respectively. Accordingly, 

the Y R image is divided into 255, which provides an 

image which can be displayed on normal display devices, 

expressed by Y d x,y. Finally, the compressed color image 

I di x,y is reproduced by Eq. (18), where denotes a gamma 

correction coefficient. In this paper, =0.5 is used 

I di x,y = I 

ix,y 

Y 

Yx,y d x,y. 18 

Figures 14–17 show some experimental results. For the next 

part, the images in (a) by the proposed model are compared 

with those in (b) by Larson’s histogram adjustment 

method. 27 In total, our results are much the same as Larson’s 

results in spite of its simple and fast algorithm. However, 

unfortunately, our result in Fig. 14 looks worse than Larson’s 

and different from other samples. It has a drawback that the 

water drops on the right side glass door are overenhanced 

thereby reducing its resolution. We have not found the cause 

of this phenomenon yet, but it may come from an improper 

choice of weights and kernel sizes to create the integrated 

surround. On the contrary, in Figs. 16 and 17, the proposed 

method could display some areas visibly which are invisible 

in Larson’s results. 28 

CONCLUSIONS 

In this paper, a concise and fast Retinex algorithm different 

from conventional MSR is proposed by integrating multiscale 

surround images into a single surround. The proposed 

model worked as well as MSR in suppressing the banding 

artifacts obtained by conventional SSR. In addition, the 

Figure 13. Histogram of luminance image by proposed Retinex of high 

dynamic range image. 

computation time was dramatically reduced by introducing 

the Gaussian pyramid. This simple model worked nicely in 

appearance improvement for both normal LDR and HDR 

images with range compression. Retinex has a goal to reproduce 

the original scene just as the observer may have seen it. 

To find the optimum parameters, we synthesized a target 

image on display visually matched to the real scene as observed 

by naked eye in the experimental room. A simple test 

target color block is captured under nonuniform illumination 

in the experimental room and used for evaluating the 

color reproducibility. Finding more robust and stable pa- 



Figure 17. Air traffic tower: a by proposed model and b by Larson 

with histogram adjustment. 

ACKNOWLEDGMENT 

The authors would like to thank Ward Larson for his help 

with the HDR images used in this paper. 

Figure 14. Bathroom: a by proposed model and b by Larson with 

histogram adjustment. 

Figure 15. Memorial Church: a by proposed model and b by Larson 

with histogram adjustment. 

Figure 16. Win office: a by proposed model and b by Larson with 

histogram adjustment. 

rameters in a full automatic mode for more complicated 

target images is left to future work involving psychophysical 

tests. 

REFERENCES 

1 J. DiCarlo and B. Wandell, “Rendering high dynamic range images”, 

Proc. SPIE 3956, 392 (2001). 

2 J. Tumblin, J. Hodgins, and B. Guenter, “Two methods for display of 

high contrast images”, ACM Trans. Graphics 18, 56 (1999). 

3 J. Tumblin and H. Rushmeier, “Tone reproduction for realistic images”, 

IEEE Comput. Graphics Appl. 13, 42 (1993). 

4 S. N. Pattanaik, J. Tumblin, H. Yee, and D. P. Greenberg, “Timedependent 

visual adaptation for fast realistic image display”, Proc. 

SIGGRAPH (ACM/Addison-Wesley, New York, 2000) pp. 47–54. 

5 K. Chiu, M. Herf, P. Shirley, S. Swamy, C. Wang, and K. Zimmerman, 

“Spatially nonuniform scaling functions for high contrast images”, 

Proceedings of Graphics Interface 93 (Morgan Kaufmann, San Francisco, 

CA, 1993) pp. 245–253. 

6 H. Kotera and M. Fujita, “Appearance improvement of color image by 

adaptive scale-gain Retinex model”, Proc. IS&T/SID 10th Color Imaging 


7 Y. Yoda and H. Kotera, “Appearance improvement of color image by 

adaptive linear Retinex model”, Proc. IS&T’s NIP21 (IS&T, Springfield, 

VA, 2004) pp. 660–663. 

8 S. Carrato, “A pseudo-Retinex approach for the visualisation of high 

dynamic range images”, Proc. 5th COST 276 Workshop (COST, 

European Science Foundation, Brussels, 2003) pp. 15–20. 

9 E. H. Land, “The Retinex”, Am. Sci. 52, 247 (1964). 

10 E. H. Land and J. J. McCann, “Lightness and the Retinex theory”, J. Opt. 

Soc. Am. 61, 1 (1971). 

11 E. H. Land, “The Retinex theory of colour vision”, Proc. R. Institution 

Gr. Britain 47, 23 (1974). 

12 E. H. Land, “Recent advances in the Retinex theory and some 

implications for cortical computations: Color vision and the natural 

image”, Proc. Natl. Acad. Sci. U.S.A. 80, 5163 (1983). 

13 E. H. Land, “An alternative technique for the computation of the 

designator in the Retinex theory of color vision”, Proc. Natl. Acad. Sci. 

U.S.A. 83, 3078 (1986). 

14 J. Frankle and J. J. McCann, “Method and apparatus for lightness 

imaging”, US Patent 4,384,336 (1983). 

15 R. Kimmel, “A variational framework for Retinex”, Int. J. Comput. Vis. 

52, 7 (2003). 

16 J. J. McCann, “Lessons learned from Mondrians applied to real images 

and color gamuts”, Proc. IS&T/SID 7th Color Imaging Conference (IS&T, 


17 J. J. McCann, “Retinex at 40”, J. Electron. Imaging 1, 6 (2004). 

18 B. Funt, V. Cardei, and K. Barnard, “Learning colour constancy”, Proc. 

IS&T/SID 4th Color Imaging Conference (IS&T, Springfield, VA, 1996) 

pp. 58–60. 

19 A. Blake, “Boundary conditions of lightness computation in Mondrian 

world”, Comput. Vis. Graph. Image Process. 32, 314 (1985). 

20 J. J. McCann and I. Sobel, “Experiments with Retinex”, HPL Color 

Summit (Hewlett Packard Laboratories, Technical Report, 1998). 

21 B. Funt, F. Ciurea, and J. McCann, “Retinex in MATLAB”, Proc. IS&T/SID 

8th Color Imaging Conference (IS&T, Springfield, VA, 2000) pp. 112–121. 

22 Z. Rahman, “Properties of a center/surround Retinex: Part 1: Signal 

processing design”, NASA Contractor Report 198194 (1995), p. 13. 



23 D. J. Jobson and G. A. Woodell, “Properties of a center/surround 

Retinex: Part 2: Surround design”, NASA Technical Memorandum 

110188 (1995), p. 15. 

24 Z. Rahman, D. J. Jobson, and G. A. Woodell, “Multiscale Retinex for 

color rendition and dynamic range compression”, Proc. SPIE 2847, 183 

(1996). 

25 D. J. Jobson, Z. Rahman, and G. A. Woodell, “Properties and 

performance of the center/surround Retinex”, IEEE Trans. Image 

Process. 6, 451 (1997). 

26 Z. Rahman, D. J. Jobson, and G. A. Woodell, “Retinex processing for 

automatic image enhancement”, Proc. SPIE 4662, 390 (2002). 

27 W. Larson, H. Rushmeier, and C. Piatko, “A visibility matching tone 

reproduction operator for high dynamic range scenes”, IEEE Trans. Vis. 

Comput. Graph. 3, 291 (1997). 

28 http://www.truview.com/images. 




Illumination-Level Adaptive Color Reproduction Method 

with Lightness Adaptation and Flare Compensation 

for Mobile Display 

Myong-Young Lee, Chang-Hwan Son and Jong-Man Kim 

School of Electrical Engineering and Computer Science, Kyungpook National University, 

1370 Sankyuk-dong, Buk-gu, Daegu 702-701, Korea 

Cheol-Hee Lee 

Major of Computer Engineering, Andong National University, 388 Seongcheon-dong, Andong, Gyeongsangbuk- 

Do 760-749, Korea 

Yeong-Ho Ha 

School of Electrical Engineering and Computer Science, Kyungpook National University, 

1370 Sankyuk-dong, Buk-gu, Daegu 702-701, Korea 

E-mail: yha@ee.knu.ac.kr 

Abstract. Mobile displays such as personal digital assistants and 

cellular phones encounter various illumination levels, different from 

the flat panel displays mainly used in indoor environment. In particular, 

in the daylight condition, the displayed images or text on a mobile 

display can be darkly perceived, which results in the degradation 

of sun readability in a mobile display. To overcome this problem, 

we proposed an illumination level adaptive color reproduction 

method with a lightness adaptation model and flare compensation. 

Lightness adaptation is a physiological mechanism to shift the photoreceptor 

response curve according to the illumination level. Thus, 

as a mobile phone is carried from an indoor to outdoor environment, 

the photoreceptor response curve automatically shifts toward a 

higher luminance to adapt to daylight intensity. Consequently, for a 

lower intensity emitted from the mobile display, the photoreceptor 

response curve becomes less sensitive, thereby decreasing the perceived 

brightness of the displayed image. Moreover, colors produced 

by mobile display can also be influenced by the flare, defined 

as ambient light reflected from the display panel, which reduces the 

maximum chroma of the mobile display gamut. Based on these 

physiological and physical phenomena, the lightness values of the 

input image are enhanced by making a linear relation between input 

luminance value estimated by device characterization and photoreceptor 

response value calculated from the lightness adaptation 

model. For the chroma component of the lightness-enhanced input 

image, chroma compensation is conducted by adding the chroma 

values of the flare multiplied by the enhancement parameter, depending 

on the hue plane of the gamut boundary. Throughout the 

experiment, the proposed algorithm not only reproduces bright and 

colorful images in the mobile display under daylight conditions, but 

also produces a solution to improve sunlight readability. 



INTRODUCTION 

Display devices such as liquid crystal displays (LCDs) and 

plasma display panels (PDPs) etc., are generally used in the 

 

IS&T Member 

Received Jun. 5, 2006; accepted for publication Oct. 2, 2006. 

1062-3701/2007/511/44/9/$20.00. 

indoor environment. As such, many display manufacturers 

have mainly focused on developing the contrast ratio, screen 

size, backlight source, and viewing angles. Even though mobile 

displays have achieved high color fidelity and good quality, 

changes in viewing conditions, i.e., the intensity or color 

temperature of the illumination considerably influences the 

original colors produced by mobile displays. Thus, viewing 

conditions have recently become a hot issue in the field of 

image quality and it has drawn considerable interest from 

display manufacturers. 1,2 One of the viewing conditions, the 

color temperature of the illumination, can make the displayed 

image appear more blue or reddish given the function 

of chromatic adaptation in a human visual system. Yet, the 

influence of color temperature is not as significant for a 

luminous body as for a reflector. In daily life, there is little 

opportunity to be in a room with incandescent or ultraviolet 

light. On the contrary, we frequently encounter various illumination 

levels between the office and the outdoor environment, 

which makes it possible to decrease the sunlight readability, 

gamut size, lightness and colorfulness of the mobile 

display. In particular, under daylight conditions, the displayed 

image on the mobile screen is perceived to be darker 

and image quality significantly deteriorates. On that account, 

various algorithms have been suggested or a new type of 

mobile display has been developed to solve this problem. 

One of the algorithms, logarithmic or power function, 

has been used to enhance the lightness of mobile phone. 3 

Since this method can simply increase the lightness of the 

displayed image, the logarithm or power curve have been 

modified based on a visual evaluation and various subject 

experiments. Yet, these have a disadvantage in that they wash 

out the color of the displayed image. Meanwhile, Monobe 

proposed a method for preserving the local contrast to 

maintain the same contrast as seen in a dark room. 4 Al- 

44

Lee et al.: Illumination-level adaptive color reproduction method with lightness adaptation and flare compensation for mobile display 

though this method can effectively preserve the whole contrast 

of an original image, the computational complexity is 

high and noise artifacts such as white points emerge in the 

detail region. There is another method to control the backlit 

unit, according to the ambient illumination levels by using a 

lux sensor. 5 This method requires a considerable amount of 

power even though a higher performance may be achieved. 

Furthermore, many display manufacturers have developed 

new types of mobile displays such as a transflective LCD to 

utilize both ambient light and backlight for displaying 

images. 6 Under dark ambient conditions, the backlight is 

turned on to illuminate, while the backlight is turned off to 

save power and utilize the ambient light under the bright 

ambient circumstance. Nevertheless, it cannot completely escape 

the influence of the daylight intensity in the outdoor 

environment. 

In this paper, we try to overcome the sunlightreadability 

problem by developing an illumination level 

adaptive reproduction algorithm and applying it to the 

transflective mobile display. The proposed method is composed 

of two steps; lightness enhancement and chroma 

compensation. To find a solution for the lightness enhancement, 

it is first analyzed why the displayed image on a mobile 

LCD is significantly perceived as dark and readability 

problems occur in daylight condition. The main cause is 

regarded as the function of the lightness adaptation in daily 

life. In general, the intensity of the daylight covers a huge 

range of about 10 8 cd/m 2 , and human eyes are capable of 

seeing about 10 5 cd/m 2 . 7,8 Nonetheless, the human eye can 

cope with a high dynamic range without much strain due to 

lightness adaptation, which is an ability to slide the photoreceptor 

response curve along the illumination level for a 

given viewing condition. Thus, as a mobile phone is carried 

outdoors, the photoreceptor response curve automatically 

adapts to the outdoor environment and becomes more sensitive 

for the daylight intensity. However, the displayed image 

is perceived as dark because the photoreceptor response 

curve becomes less sensitive to the lower intensity emitted 

from the mobile display. Based on this kind of physiological 

mechanism, lightness enhancement is proposed by conducting 

a linearization process between the input luminance and 

photoreceptor response to obtain a smooth tone reproduction. 

However, after doing the lightness enhancement, satisfactory 

results cannot be obtained because lightness enhancement 

only washes out the color of the displayed image. 

Moreover, the flare, some of ambient light that is reflected to 

the front glass plate of the display, physically decreases the 

color gamut through desaturation. 9,10 Accordingly, in this 

paper, chroma compensation will be considered together 

with the lightness enhancement to obtain a better displayed 

image on the mobile display 

The remainder of this paper is organized as follows. The 

following section provides an outline of the proposed algorithm, 

followed by detailed explanations of the proposed 

method consisting of four subsections, i.e., Flare Calculation, 

Lightness Enhancement, Chroma Compensation, and 

the Construction of the Three-dimensional (3D) Lookup 

Table. In the Flare Calculation subsection the physical effect 

of flare will be investigated to determine the changes in the 

mobile gamut, and flare estimation will be described based 

on the CIE 122-1966. In the Lightness Enhancement and 

Chroma Compensation subsections the main cause of deteriorating 

sunlight readability will be analyzed based on the 

human visual system and the illumination level adaptive 

color reproduction method will be proposed. Subsequently, 

a method for the design of the 3D lookup table will be 

explained briefly in the next subsection. In the Experiments 

and Results section, subjective experiments will be conducted 

under daylight condition, and the performance of 

various algorithms will be compared and analyzed using 

z-score evaluation. From these results, the conclusions will 

be presented in the final section. 

PROPOSED METHOD 

Figure 1 shows the flowchart of the proposed algorithm that 

achieves illumination level adaptive color reproduction. 

First, the TSL 2550 lux sensor is built into a mobile phone to 

detect ambient light intensity. According to the measured 

intensity level, the amount of flare expressed as the CIEXYZ 

value is calculated on the basis of CIE 122-1966, which is 

added to the CIEXYZ values of the original image estimated 

by using a conventional monitor characterization such as the 

gain offset gamma (GOG) model, S-curve model, or piecewise 

linear interpolation. Then, for luminance component of 

CIXYZ values, lightness enhancement is implemented by establishing 

a linear relationship between the luminance values 

and the cone response values to obtain perceived tone reproduction, 

where the cone response values corresponding to 

the luminance value are simply calculated from the lightness 

adaptation model. Following the lightness enhancement, the 

gamut boundary description was established by the mountain 

range segment method and chroma compensation was 

successively executed by adding the chroma values reduced 

by the flare to those of original image, yielding a colorful 

image. 11 However, since this kind of serial-based procedure 

Figure 1. Flowchart for the proposed algorithm. 



is not appropriate for real-time processing, a lookup table 

representing daylight intensity is designed based on the 

sampled RGB data. 

Flare Calculation 

Before calculating the amount of the flare, mobile LCD characterization 

is performed by piecewise linear interpolation to 

establish a relationship between the RGB values and tristimulus 

values (CIEXYZ or CIELAB). Model-based characterization, 

such as the GOG or S-curve models, is not well 

suited for a mobile phone because of the behavior imposed 

by the system design. 12 In the CIE 122-1996, flare is defined 

as the portion of the ambient light reflected from the display 

panel and is added to the colors produced by the mobile 

LCD 10 

X 

Y =X 

Y +X 

Y . 1 

ZDisplay 

ZLCD 

ZFlare 

Color appearance on a mobile LCD is very much affected by 

ambient lighting, since the human visual system changes its 

sensitivity according to the surroundings. However, the colors 

produced by a mobile LCD are physically affected by 

ambient light. When ambient light illuminates a mobile 

LCD, the LCD screen reflects some of this light. This reflection 

is added to the colors that are produced by the mobile 

LCD. The amount of the flare is expressed as 

X 

Y = R · M 

x Ambient 

1 

y 

ZFlare 

Ambient 

y Ambient 

1−x Ambient − y Ambient, 

where R is the reflection ratio of the display screen and 

x Ambient ,y Ambient is the chromatic diagram of the ambient 

light; M is the intensity of the ambient light (lux) taken from 

the TSL2550 lux-sensor. To estimate the reflection ratio of 

the mobile LCD, the CIEXYZ values of the black patch are 

measured using a colorimeter in a dark room and in the 

outdoor environment. The amount of flare in Eq. (2) is then 

obtained by calculating the difference for each measured 

CIEXYZ value, and the x Ambient ,y Ambient is given as D65 

(0.3127, 0.3290). By substituting these values into Eq. (2), 

the reflection ratio is acquired as seen in Table I. The results 

show that the reflection ratio for a mobile LCD is generally 

between 0.5% and 2%, and it is lower than that of the cathode 

ray tube (CRT) monitor which is between 3% and 5%. 

From the reflection ratio, the gamut of the mobile LCD is 

investigated as to how the flare influences the gamut size of 

a mobile LCD. Figure 2 shows the gamuts that correspond to 

daylight amount of 5000 and 10 000 lux, compared with the 

gamut measured in a dark room. As the level of daylight 

increases from 5000 to 10 000 lux, it can be observed that 

the chroma values decrease depending on the hue plane, 

while the lightness values increase. 

2 

Table I. Measured black patch and estimated reflection ratio. 

X Y Z R 

0 lux 0.52 0.47 0.77 

500 lux 1.78 1.91 2.63 0.008 

4000 lux 12.76 13.5 14.63 0.01 

9000 lux 29.20 30.4 39.73 0.01 

15 000 lux 47.92 49.5 59.7 0.011 

Lightness Enhancement Method Based on the Lightness- 

Adaptation Model 

One of the problems of mobile LCDs is that displayed images 

are perceived as dark under the outdoor environment 

due to lightness adaptation. Lightness adaptation is a physiological 

mechanism to displace the visual response curve 

according to the ambient level, analogous to automatic exposure 

control in a digital camera. Figure 3 shows visual 

response shifting to adapt to ambient intensity, and it illustrates 

why the displayed image is perceived dark in the outdoor 

environment. In Fig. 3, if the indoor environment 

200 cd/m 2 changes to an outdoor environment 

2000 cd/m 2 , the visual response curve shifts toward a 

higher luminance to adapt ambient level, i.e., automatic 

HDR function. However, the maximum luminance of the 

mobile LCD is limited to about 100 cd/m 2 . Thus, it can be 

observed that a relative cone response of 0.6 under indoor 

environment is reduced to 0.22 at the maximum luminance. 

This is why the quality of the displayed image or text in a 

mobile phone significantly deteriorates in the outdoor environment. 

Based on this kind of physiological mechanism, lightness 

enhancement is carried out by following the procedure 

in Fig. 4. An input RGB value is converted into a CIEXYZ 

value by using the piecewise linear interpolation. Lightness 

enhancement is executed only for the luminance component 

of the XYZ value, while the remainder of the components 

one left intact. First, the flare is added with an input luminance 

value, which is then mapped to a cone response by 

using the lightness adaptation model 

Y = Y image + Y flare , 

Y n 

3 

R cone = fY = = I 

Y n + n A , 4 

where Y image and Y flare are the luminance values of the input 

image and flare, respectively. In general, the parameters 

,,n are variables, not constant values. However, since 

the range of parameters is extensive, it is necessary to fix the 

values of the parameters to simulate the lightness adaptation 

model. In Ref. 7, Ledda suggested a method to compute the 

localized adaptation intensity M in Eq. (5) and to set the 



Figure 2. Comparison of the gamut under an outdoor environment solid frame and an indoor environment 

wire frame: a 5000 lux side, b 10 000 lux side, c 5000 lux top, andd 10 000 lux top. 

Figure 4. Procedure for lightness enhancement using the lightness adaptation 

model. 

I A = M , 

assumption of Lamberitian reflection cone response. The sampled input luminance values 

5 

where M is the ambient intensity (lux) acquired by the lux 

Figure 3. Cone response curve according to the intensity of the ambient 

light. 

sensor. 

Second, the corresponding luminance Y for the cone 

response R cone is found through linearization of the input 

luminance Y to establish a linear relation between input 

range of the parameters’ values ,. Therefore, we 

adapted the parameter , values and set the range of 

n-accuphy to the lightness enhancement experiment, which 

will be referred to at the end of this section; is the halfsaturation 

parameter, and I A is the adaptation level calculated 

by dividing the ambient intensity (lux) with on the 

luminance and cone response for the lightness enhancement. 

Linearized cone response can be acquired by exchanging the 

cone response with the input luminance using a piecewise 

linear interpolation because the inverse cone response curve 

in Eq. (4) is not directly calculated. 14 Figure 5 shows the 

general linearization method used to calculate the inverse 



y 0 ,y 1 ,...,y n are transformed to a cone response value 

R cone,0 ,R cone,1 ,...,R cone,n using Eq. (4). These cone response 

values are normalized to an amount of one and are 

stored in one-dimensional (1D) lookup table (LUT). For an 

arbitrary input luminance value, piecewise linear interpolation 

is applied to the 1D lookup table, thus creating the 

output cone response curve in Fig. 5(a). Then, inverse cone 

response curve in Fig. 5(b) is simply obtained by switching 

the cone response value with the luminance value stored in 

the 1D LUT. Therefore, a new input value R cone for the 

inverse cone response can be calculated as follows: 

R cone = Y max − Y min 

R max − R max R cone − R min , 6 

where Y max and Y min are the maximum and minimum luminance 

values, respectively while R max and R min are maximum 

and minimum cone response values. 

Finally, the corresponding luminance Y for the input 

value R cone is obtained by applying the piecewise linear 

interpolation to the 1D LUT in Fig. 5(b). This value is then 

combined with the intact color components and is transformed 

into the CIELCH color space for the subsequent 

application of the chroma compensation. 2 At this point, to 

convert the CIEXYZ values into CIELCH values, the reference 

CIEXYZ value is defined as the amount of ambient 

light that represents a white object in the scene. On the other 

hand, the result of the proposed lightness enhancement depends 

on the values of parameter n. Thus, to find the appropriate 

parameter value, the observer should select the 

best results of the lightness enhanced images under the outdoor 

environment. Table II shows the appropriate parameter 

values corresponding to daylight intensity. From the subjective 

experiment, the parameter value becomes higher as the 

daylight intensity rises; because a large value of the parameter 

increases the degree of the lightness enhancement. Figure 

6 shows the cone response curve according to the parameter 

values for 1000 and 10 000 lux. 

Chroma Compensation Using the Flare 

When only lightness enhancement is applied to the input 

image, the color of the enhanced image is washed out due to 

the influence of the flare. Therefore, to compensate for the 

reduced chroma physically, the chroma difference between 

two types of environment, i.e., darkroom and outdoors, is 

added to the CIELCH value acquired from lightness enhancement 

as shown in Fig. 7. However, since the chroma 

difference depends on the hue value as seen in Fig. 4, 

chroma compensation should be applied considering each 

hue value individually 

C diff = C − C flare , 

C = C + · C diff , 

7 

where C and C flare are the chroma values from the darkroom 

and outdoor environment, respectively. C diff is the chroma 

difference between C and C flare . The compensated chroma 

value C is adjusted according to with the enhancement 

parameter . If the chroma value of input image C is close 

to the gamut boundary of mobile display and is added with 

Table II. Appropriate parameter values according to the daylight intensity. 

Figure 5. Linearization method: a construction of the cone response 

curve using piecewise linear interpolation and b construction of the linearized 

cone response curve using piecewise linear interpolation. 

Lux 1000 5000 10 000 20 000 

n 2.0 2.0 2.5 3.5 



Figure 8. Chroma compression around the gamut boundary. 

Figure 6. Cone response curve according to various parameter values 

for 1000 and 10 000 lux. 

Figure 7. Concept of chroma compensation based on chroma 

difference. 

C diff , the compensated chroma value C can get outside the 

gamut boundary that the mobile display is capable of reproducing. 

Thus, the enhancement parameter is modified in 

consideration of the gamut boundary, as seen in Fig. 8 

if C C gamut − C diff 

=1 

C gamut − C , 8 

, otherwise 

C diff 

where is the compression starting point parameter and 

C gamut is the gamut boundary calculated by using the mountain 

range method developed by Braun and Fairchild. 11 This 

method uses griddling and interpolation to arrive at a data 

structure consisting of a uniform grid in terms of lightness 

and hue, and it stores the gamut’s most extreme chroma 

values for each of the grid points. The boundary value has 

101 and 360 levels for each grid points. If the input chroma 

value is inside C gamut −·C diff , the chroma difference is 

added to the input chroma value without compression. Otherwise, 

compression compensation is executed by using the 

compression starting point parameter , which can be set 

flexibly values of 1.0, 1.5, and 2.0 are used in this paper. If 

is over 2.0, chroma compensation is not effective through 

the experiment, while a clipping artifact is generated if the 

value is less than 1.0. 

The Construction of the 3D Lookup Table 

A 3D LUT is constructed to represent the intensity of daylight 

(10 000 lux) for real-time processing. The input RGB 

digital values are uniformly sampled by the nnn grid 

points, which are processed by the proposed algorithm, thus 

resulting in the output RGB values. The sampled input and 

output RGB digital values are stored in the 3D LUT and 3D 

interpolation such as trilinear, pyramid, or tetrahedral is 

used to calculate the output RGB values for the arbitrary 

input RGB values. 14 This 3D LUT can be inserted into the 

mobile phone and functions well in a mobile environment 

without the difficulties associated with the memory and 

computation. 

EXPERIMENTS AND RESULTS 

To test the sunlight readability of mobile display for various 

methods, Transflective PDA (SPH-M4000) made by Samsung 

Electronics was used as the testing device, and ten observers 

consisting of five ordinary citizens and five color imaging 

experts participated in the subjective experiment. The 

average age of observers is 29 years old; ages range from 

27 to 31 years old and one observer is female. In addition, 

to ensure the changeable viewing conditions of the real 



world, we use lighting equipment supported by Samsung 

Electronics to control the intensity of illumination from 0 to 

20 000 lux. Thus, the subjective experiment is conducted in a 

dark room using this lighting equipment for two light conditions, 

i.e., 2000 and 10 000 lux to represent cloudy and 

bright days, respectively. Figure 9 shows the original images, 

and Figs. 10 and 11 show the enhanced test images to be 

displayed on the personal digital assistant under two lighting 

conditions. Figure 10(a) shows the resulting image when the 

logarithmic function is used. Although this method increases 

the amount of lightness in the original image, the 

color of the original image is washed out, and thus, colorfulness 

is considerably decreased under these conditions. 

Figure 10(b) shows the resulting image when using 

Monobe’s method which preserves the local contrast. This 

method may maintain a contrast ratio similar to the original 

image seen in a darkroom. However, noise artifacts like 

white points appear in the leaf regions due to excessive contrast 

enhancement. In addition, this method has the complex 

computations that are not suitable for the implementation 

of real-time processing. Figure 10(c)–10(e) show the 

resulting images of the proposed methods with different 

values. In Figs. 10(c)–10(e), it is seen that the colorfulness of 

the resulting images is significantly enhanced, which improves 

sunlight readability and provides pleasure to the observers 

under the two lighting conditions, especially at 

10 000 lux. Also, the chroma values of the resulting images 

are perceived similar to those of the original image in a dark 

room, even though the chrome values are excessively enhanced. 

Figure 11 shows other resulting images with five 

enhancement methods, and we can find the same effect for 

Figure 10. Enhanced park images under 10 000 lux: a lightness enhancement 

using the logarithmic function, b Monobe’s method preserving 

the local contrast, c the proposed method with =2.0, d the 

proposed method with =1.5, and e the proposed method with 

=1.0. 

Figure 9. Test images: a park, b cap, c girl, and d woman. 

Figure 11. Enhanced woman images under 10 000 lux: a lightness 

enhancement using the logarithmic function, b Monobe’s method preserving 

the local contrast, c the proposed method with =2.0, d the 

proposed method with =1.5, and e the proposed method with 

=1.0. 



Table III. Z-score values of ordinary citizens: a 2000 and b 10 000 lux. 

a Daylight condition: 2000 lux 

Image L M LC 

= 2.0 

LC 

= 1.5 

LC 

= 1.0 

Park −12.01 

−12.53 

−2.81 

−2.29 

−3.24 

−3.24 

4.94 

4.94 

13.12 

13.12 

Cap −13.12 

−13.12 

−0.52 

−3.83 

−0.59 

−1.11 

4.35 

1.7 

9.88 

16.36 

Woman −16.36 

−16.36 

−8.18 

−8.18 

0 

3.24 

8.18 

4.94 

16.36 

16.36 

Girl −13.12 

−13.12 

−1.7 

−4.35 

7.07 

3.83 

5.53 

12.01 

2.22 

1.63 

b Daylight condition: 10 000 lux 

Image L M LC 

= 2.0 

LC 

= 1.5 

LC 

= 1.0 

Park −13.12 

−13.12 

−1.7 

−1.7 

−3.24 

0 

4.94 

1.7 

13.12 

13.12 

Cap −16.36 

−16.36 

−1.7 

−4.94 

−3.24 

−3.24 

4.94 

8.18 

16.36 

16.36 

Woman −16.36 

−16.36 

−1.7 

−8.18 

0 

0 

4.94 

8.18 

13.12 

16.36 

Girl −16.36 

−16.36 

−8.18 

−8.18 

0 

0 

8.18 

8.18 

16.36 

16.36 

Table IV. Z-score values of color imaging experts: a2000 and b 10 000 lux. 

a Daylight condition: 2000 lux 

Image L M LC 

= 2.0 

LC 

= 1.5 

LC 

= 1.0 

Park −12.01 

−12.01 

−9.29 

−12.53 

−3.24 

0 

8.18 

8.18 

16.36 

16.36 

Cap −16.36 

−16.36 

8.18 

4.35 

−4.94 

−1.7 

0 

0.59 

13.12 

13.12 

Woman −13.12 

−16.36 

−8.18 

−8.18 

3.24 

0 

4.94 

8.18 

13.12 

16.36 

Girl −13.12 

−12.53 

−1.11 

−4.35 

0.59 

3.83 

12.01 

12.01 

1.63 

1.04 

b Daylight condition: 10 000 lux 

Image L M LC 

= 2.0 

LC 

= 1.5 

LC 

= 1.0 

Park −12.01 

−12.01 

−2.81 

−2.81 

−3.24 

−3.24 

4.94 

4.94 

13.12 

13.12 

Cap −16.36 

−16.36 

4.94 

4.94 

−4.94 

−4.94 

3.83 

3.83 

12.53 

12.53 

Woman −16.36 

−16.36 

−8.18 

−8.18 

0 

0 

8.18 

8.18 

16.36 

16.36 

Girl −16.36 

−16.36 

−8.18 

−8.18 

0 

0 

8.18 

8.18 

16.36 

16.36 



each method, with the exception that Monobe’s method 

does not uphold any noise artifacts applied to this image. 

To conduct subjective evaluation, four test images and 

five enhanced methods are used and their paired images are 

randomly selected to obtain z-score values. 15 Tables III and 

IV show the z-score evaluations of ordinary citizens and 

color imaging experts for two lighting conditions, where L, 

M, LC=2.0, LC=1.5, and LC=1.0 represent the 

lightness enhancement using logarithmic function, Monebe’s 

method preserving the local contrast, proposed methods 

with three different values, respectively. The numbers in 

parentheses represent the z-score values obtained by the second 

experiment under equivalent conditions. In Tables III 

and IV, three differences in the z-score value obtained by five 

observers can be regarded as the same results because the 

frequency is almost equal which indicates that the ith 

method is judged better than the jth method. Thus, the results 

of z-score values are almost the same at the 10 000 lux, 

irrespective of observer type and repeated experiment. However, 

small differences occur in the 2000 lux condition, depending 

on observer type. For the “park” image, ordinary 

citizens prefer the image resulting from of Monobe’s method 

more than that of the lightness enhancement method, while 

color imaging experts give better marks to the lightness enhancement 

method because of noise artifacts like white 

points in the leaf region. Similarly, for the “cap” image, the 

z-score value of the proposed method is lower than that of 

Monobe’s method due to the sharpness problem. From these 

results, it is seen that the color imaging experts attach importance 

to image quality such as noise and sharpness, relative 

to a slight increase in readability. In addition, as the 

intensity of illumination changes from 2000 to 10000 lux, we 

found that in the “girl” image, the z-score value of the proposed 

method with =1.0 increases considerably. The reason 

is that the fine clipping artifact in the cloth region is 

indistinguishable due to the influence of higher illumination 

level. Consequently, the proposed method with =1.0 has 

the best performance among the five methods, and we found 

that the noise or clipping artifact is an important factor to 

influence the z-score evaluation depending on observer type 

and illumination level. However, the results of z-score evaluation 

are almost the same irrespective of a number of experiments 

conducted. 

CONCLUSION 

This paper suggests and analyzes problems that can occur 

for the mobile display in an outdoor environment as a result 

of human lightness adaptation and flare phenomena. First, 

we explained why readability or image quality of mobile 

phones is significantly degraded under daylight condition 

based on lightness adaptation. The lightness enhancement 

algorithm is then proposed to increase the luminance of the 

input RGB image by the linearization process between the 

input luminance and cone response. Second, the influence of 

the flare is investigated to determine the variations of the 

mobile gamut, and it can be observed that the maximum 

chroma values are reduced differently depending on the hue 

plane. From this observation, chroma compensation is executed 

by adding the differentially reduced chroma values 

according to the hue plane with lightness enhanced input 

image. Finally, a 3D lookup table, composed of RGB grid 

points, is implemented to achieve real-time processing. The 

experiment shows that the lightness enhancement and 

chroma compensation algorithm is well suited for mobile 

LCDs, thus reproducing more colorful and brighter results 

in the outdoor environment. Furthermore, we expect that 

the proposed algorithm can be applied to other portable 

devices. 


This work is financially supported by the Ministry of Education 

and Human Resources Development (MOE), the 

Ministry of Commerce, Industry and Energy (MOCIE), and 

the Ministry of Labor (MOLAB) through the fostering 

project of the Lab of Excellency. 

REFERENCES 

1 N. Mornoney, M. D. Fairchild, R. W. G. Hunt, C. Li, M. R. Luo, and T. 

Newman, “The CIECAM02 color appearance model”, Proc. IS&T/SID 

10th Color Imaging Conference (IS&T, Springfield, VA, 2002) pp. 23–27. 

2 M. D. Fairchild, Color Appearance Models (Wiley, New York, 2005). 

3 F. Drago, K. Myszkowski, T. Annen, and N. Chiba, “Adaptive 

logarithmic mapping for displaying high contrast scenes”, 

EUROGRAPHICS 2003 (2003). 

4 Y. Monobe, H. Yamashita, T. Kurosawa, and H. Kotera, “Fadeless image 

projection preserving local contrast under ambient light”, Proc. IS&T/ 

SID 12th Color Imaging Conference (IS&T, Springfield, VA, 2004) pp. 

130–135. 

5 S. H. Kim, “Device and method for controlling LCD backlight”, US 

Patent 6,812,649 B2 (2004). 

6 X. Zhu, Z. Ge, T. X. Wu, and S. T. Wu, “Transflective liquid crystal 

displays”, IEEE/OSA J. Display Technol. 1 (2005). 

7 P. Ledda, L. P. Santos, and A. Chalmers, “A local model of eye adaptation 

for high dynamic range images”, Proceedings of the 3rd International 

Conference on Computer Graphics, Virtual Reality, Visualization and 

Interaction in Africa (ACM Press, New York, 2004) pp. 151–160. 

8 E. Reinhard and K. Devlin, “Dynamic range reduction inspired by 

photoreceptor physiology”, IEEE Trans. Vis. Comput. Graph. 11, 13–24 

(2005). 

9 J. Laine and M. Kojo, “Illumination-adaptive control of color 

appearance: a multimedia home platform application”, Research Report 

TTE4-2004-4, VTT Information Technology (Jan. 2004). 

10 N. Katoh and T. Deguchi, “Reconsideration of CRT monitor 

characteristics”, Proc. IS&T/SID 5th Color Imaging Conference (IS&T, 


11 G. J. Braun and M. D. Fairchild, “Techniques for gamut surface 

definition and visualization”, Proc. IS&T/SID 5th Color Imaging 


12 G. Sharma, “LCDs versus CRTs color calibration and gamut 

considerations”, Proc. IEEE 90 (2002). 

13 N. Morony and P. Alto, “Usage guidelines for CIECAM97s”, Proc. IS&T 

PICS Conference (IS&T, Springfield, VA, 2000) pp. 164–168. 

14 H. R. Kang, Color Technology for Electronic Imaging Devices (SPIE 

Optical Engineering, Bellingham, WA, 1996). 

15 T. C. Hseue, Y. C. Shen, P. C. Chen, W. H. Hsu, and Y. T. Liu, 

“Cross-media performance evaluation of color models for unequal 

luminance levels and dim surround”, Color Res. Appl. 23, 169–177 

(1998). 




Influence of Paper on Colorimetric Properties of an Ink 

Jet Print 

Marjeta Černi~ 

Pulp and Paper Institute Ljubljana, Bogiši~eva 8, 1000 Ljubljana, Slovenia 

E-mail: meta.cernic@icp-lj.si 

Sabina Bra~ko 

Faculty of Natural Sciences and Engineering, University of Ljubljana, Snežniška 5, 1000 Ljubljana, Slovenia 

Abstract. Paper for ink jet printing has to obtain optimal printing 

runnability, printability, and printing quality. Therefore, it must have 

some specific properties that ensure optimal drying time, mechanical 

stability of a print, and its light and water resistance. The paper 

surface should enable the printing ink to be dried as fast as possible. 

The aim of the applied research was to determine how an ink jet 

color print on paper changes with time immediately after printing, 

and how long it takes for the color print to stabilize. Color differences 

* 

E ab 

were measured that appeared on print after a certain amount 

of time with regards to values attained immediately after printing. 

The influence of paper on colorimetric properties and optical density 

of a print was analyzed by measuring some structural, surface, and 

sorption properties. The values attained show that the paper surface 

should enable wetting and ink penetration in paper structure. The 

biggest changes in colorimetric properties of the print became visible 

during 1 h after printing; however, color print finally stabilizes 

only after 96 h. Research results confirmed the importance of paper 

sorption properties for obtaining high-quality ink jet color prints. 



Received Mar. 16, 2006; accepted for publication Aug. 18, 2006. 

1062-3701/2007/511/53/8/$20.00. 

INTRODUCTION 

The important parameters for producing a quality print are 

the properties of inks, printers, and paper. Quality paper 

should enable images of high contrasts and excellent reproduction 

of lively colors and sharp outlines. In order to 

achieve these properties, a print needs to be dried carefully 

since droplets should not spread on the surface. To obtain 

proper print quality, it is important to have a thorough 

knowledge of paper characteristics and ink properties. The 

properties of a print under particular printing conditions in 

conventional printing techniques are well defined by a number 

of standards. Although producers of ink jet printers refer 

to certain recommendations for inks and paper, proper standards 

have not been developed or published yet. 1–5 

Paper for ink jet printing must have some specific properties 

that ensure optimal drying time, mechanical stability 

of a print, and its resistance to light and water. 6–9 The paper 

surface should enable the printing ink to be dried as fast as 

possible by absorption, adsorption, and evaporation, which 

depend on sorptive properties of paper and climatic conditions 

in a certain space. 10–12 According to the ink structure 

the ink jet printer can be divided into two categories, i.e., 

water-based ink jet and phase change ink jet, which is generally 

more substrate independent. Immediately after the 

water-based ink contacts the paper, all the interactions between 

ink droplet and paper take place. At the same time, 

the ink drying process begins and progresses until the ink is 

immobilized on the paper. The way the ink dries on the 

paper can be very critical to the quality of the final printed 

image, because the deformation of the print image, such as 

feathering, wicking, paper expansion, and bleed-through, all 

take place before the ink is immobilized. 13–15 

The ink drying process involves three major routes that 

govern the quality of print image that describes Fig. 1: 

evaporation of ink carrier (water or solvent), XY-direction 

spreading (ink traveling on paper surface), and Z-direction 

penetration (ink absorbed into paper). The ink that dries 

quickly through evaporation generally offers less time for ink 

spread and results in sharp and less deformed image. Feathering 

and wicking of the printed image are generally results 

of extensive XY-direction spreading. Bleed-through, as well 

as the optical density of the image, is strongly affected by the 

depth of Z-direction penetration. The speed of ink traveling 

through each route is generally different and depends on the 

ink formulation as well as the chemical and physical structures 

of paper. This is why we can see differences in print 

quality on different papers, and different inks offer different 

print quality on the same paper. In most cases, the rate of 

evaporation is much slower than the rate of XY-spreading 

and Z-penetration, which become the primary actions that 

contribute to the final shape of the printed image. 15 The 

ideal “Case B” for the ink to dry on the paper (Fig. 1) is to 

control carefully the rate of both XY-spreading and 

Z-penetration via adjusting the chemical and physical compositions 

of the paper. If both XY-spreading and 

Z-penetration take place in a desired way, the printed image 

should be able to expand proportionally to the original print 

out from the printing head. The resolution as well as the 

optical density of the print image will be retained and the 

feathering and wicking, as well as the bleed-trough problems, 

should be reduced. 16,17 

53

Černi~ and Bra~ko: Influence of paper on colorimetric properties of an ink jet print 

Table I. Paper properties. 

Properties Paper 1 Paper 2 Paper 3 

Grammage, g/m 2 80.7 79.1 85.0 

Specific volume, cm 3 /g 1.26 1.29 1.27 

Formation index, M/K 3-D, 49.3 36.6 51.9 

Ash content, % 

• 500 °C 24.1 19.3 11.0 

• 900 °C 14.0 11.4 10.0 

Smoothness, Bekk, s 

Figure 1. Three cases of drying mechanisms of water-based ink-jet drop 

on a plain paper. 15 

• Top side A 14 19 12 

• Bottom side B 16 20 16 

Porosity, Gurley, s 11 23 36 

EXPERIMENTAL 

The goal of this research study was to investigate how the 

sorption properties of paper affect both the XY-spreading 

and Z-penetration during the ink drying process. Therefore, 

the changes in colorimetric properties that occur on prints 

were observed in a defined time interval. In addition, it was 

important to define the amount of time after which the 

measured values are stabilized. The influence of paper on 

colorimetric properties and optical density of a print was 

analyzed by measuring structural, surface, and sorption 

properties. 18–20 

Materials and Methods 

Paper Properties 

Three paper grades for ink jet printing weighing 80 g/m 2 

made by different producers were used. Paper samples 1 and 

2 (MOTIF office paper and Rotokop Radeče, respectively) 

were of regular quality whereas sample 3 (Epson ink jet paper) 

was lightly surface coated and thus intended for high 

quality prints such as photographs. A comparative analysis 

of physical-chemical and surface properties was conducted 

in order to determine the influence of paper structure on the 

change of colorimetric properties of an inkjet print. Samples 

1, 2, and 3 were tested under standard climate conditions 

(ISO 187). The following analyses were performed on the 

basis of standard or nonstandard testing methods: 

Water absorptivity, Cobb 60, g/m 2 

• Top side A 24 20 37 

• Bottom side B 23 21 29 

Contact angle, FibroDat, 2 s/50 s, ° 

• Top side A 94/85 107/102 11/0 

• Bottom side B 91/91 97/85 105/84 

Surface tension, FibroDat, side A, mN/m 

• Total 96 98 70 

• Dispersion part 92 90 24 

• Polar part 4 8 46 

• Basic physical properties: grammage (ISO 536), thickness, 

specific volume (ISO 535), ash content (ISO 2144) 

• Paper homogeneity: formation index—Kalmes M/K 

3-D (Pulp and Paper Institute method). 

• Surface properties: Bekk smoothness (ISO 5626), Gurley 

porosity (ISO 5636-5). 

• Sorption properties: Cobb 60 water absorption (ISO 

535), contact angle (TAPPI 458) and surface tension— 

DAT 1100 (Fibro System AB). 

The results of tested properties of papers 1, 2, and 3 are 

shown in Table I and Fig. 2. 

Figure 2. Contact angle for papers 1, 2, and 3 in dependence of time. 



The Monitoring of Colorimetric Properties of Prints 

The 33 cmcolor testing chart with CMYK color fields of 

100% and 50% color application intensity was created by 

means of the Adobe Photoshop image software. Color and 

black cartridges (Epson color ink for the Epson Stylus- 

Color 900 printers for 1400 dpi resolution with A4, 

PHOTO, and color print settings) were used for printing. 

The L * a * b * values of color print samples on individual paper 

grade were measured according to the ISO 13656 standard 21 

by means of the spectrophotometer Spectrolino (Gretag 

Macbeth, D50/2° lighting, 45°/0 measurement geometry, a 

4mmmeasuring aperture, on black basis) in defined time 

intervals in order to determine how long it does take for the 

color to dry and thus for its colorimetric properties to stabilize. 

The measurements were divided into two groups: 

first, color differences were monitored in shorter time intervals 

after 3, 6 10, 15, 20, 30, 60, 90, and 120 min. Inthe 

second case, color differences were measured after longer 

periods, that is, after 1, 2, 3, 4, and 7 days. Measurements 

were conducted at constant temperature of 21 °C±2 °C and 

relative humidity of 32% ±2%. Between each measurement, 

samples were kept in the dark. The tested values were compared 

with results obtained immediately after printing. 

* 

The calculated color difference E ab was monitored and 

inserted into diagrams separately for each paper sample and 

each CMYK color. 22,23 * 

Color differences E ab for CMYK 

color samples are represented in Figs. 3 and 4. 

Colorimetric Properties of Dry Prints 

Optical density of the color print is usually the only parameter, 

which is measured during the printing process. 24 Optical 

density of a dry print (D) was measured by the densitometer 

RD 918 (Gretag Macbeth), 7 days after printing. The 

results are presented in Fig. 5. Figure 6 represents the color 

* 

differences E ab between the dry prints (7 days after printing) 

and the prints immediately after printing. 

Cross-Section of Color Prints 

Qualitative microscopic analysis of color print cross sections 

was made by cryoscopic microtome at −25 V and minutely 

examined, under an optical microscope at a magnification of 

160. The results for black print are presented in Fig. 7. 

RESULTS AND DISCUSSION 

Characterization of Paper Substrates 

Table I summarizes the mean values of basis structural, surface, 

and sorptive properties of papers 1, 2, and 3. A visual 

analysis of paper samples proved papers 1 and 2 to be natural 

and surface nontreated—they are not visibly two sided. 

However, paper 3 is surface pigmented on the topside 

whereas the bottom side is similar to the other two samples. 

By all paper samples 10% to 14% ash content has been 

obtained. The high values of ash content at 500 °C are due 

to calcium carbonate being used as paper filler in all samples 

except sample 3, which is filled by clay or other pigments on 

silicate basis proven by a very small change in ash content at 

different temperatures. The topside of paper 3 is very white, 

whereas the bottom side is slightly yellow. Two-sidedness is 

thus apparent. 

Paper homogeneity or formation is defined by transmission 

of light through paper, which can be either satisfactory 

or unsatisfactory depending on the appearance of the surface 

in transmitted light. 25 The test was conducted on an M/K 

3-D analyzer as transmission of light through an A4 paper 

sheet. Homogeneity of paper on the basis of relative weight 

deviation of a certain spot in comparison to average weight 

is defined by means of measuring both the size of distributed 

flocks and empty spots, as well as by measuring formation 

using the FI-formation index. Growing deviations in 

relative weight decrease the level of formation index and 

thus the quality of homogeneity. Based on practical experience, 

the minimum values of FI should be around 30 in 

black and white printing and more than 50 in color 

printing. 6,8,13 

Since all paper samples achieved Bekk smoothness values 

in the range of 12 to 20 s they could be classified as machine 

calendered papers, which are not appropriate for products of 

high printing quality. Slight two-sidedness was observed 

with all paper samples. 

The results of testing air porosity by Gurley show slight 

differences between the papers 1, 2, and 3. A very high porosity 

was achieved by paper 1 11 s whereas paper 2 obtained 

a slightly lower porosity value 23 s. All values obtained 

correspond to requirements for printing runnability 

in electrophotographic printing and are most probably appropriate 

for satisfactory runnability in ink jet printing as 

well. 10,11 

All paper samples obtained surface absorptivity of water 

by Cobb-60 values of 20 g/m 2 or higher, which point to a 

lower quality of sizing (Table I). The values are appropriate 

for offset printing but not for electrophotography, which 

requires a nonabsorbent surface with Cobb values lower 

than 20 g/m 2 . According to practical experience, the obtained 

values are most probably appropriate for ink jet printing. 

A considerable deviation was observed in paper 3, which 

exhibited high absorptivity on the top, pigmented side 

(Cobb values is 35 g/m 2 —denoting low-quality sized paper) 

and a slightly lower level of absorptivity on the bottom side. 

The dynamic wetting interaction between a liquid and a 

paper surface can reveal problems affecting printing, sizing, 

or coating. 13–15 Dynamic contact angle and spreading rate of 

ink drops were measured from side images of the drop profile, 

which was monitored as a function of time with a DAT 

1100 instrument (Fibro Systems AB), with a time resolution 

of 20 ms. The measurements of FibroDat contact angle (Fig. 

2) led to a similar conclusion as measurements of Cobb 

values. After 2s, the measured contact angles of samples 1 

and 2 are 90° and 105°, respectively, which denote higher 

hydrophobicity. After 50 s, these values lowered by 5° to 10°. 

The topside of paper 3 achieved total water absorption 

within 2s. Its surface is thus completely water absorbent, 

which is probably caused by a specialty pigment coated surface, 

which enables optimal absorption of inks in ink jet 

printing. The results of surface absorptivity show that paper 



Figure 3. Color difference E * ab of CMYK prints on papers 1, 2, and 3, 

up to 120 min after printing. 

Figure 4. Color difference E * ab of CMYK prints on papers 1, 2, and 3, 

up to 7 days after printing. 

3 is not appropriate for offset and electrophotographic 

printing. 

The drying phenomena of water-based ink on paper can 

be influenced by several parameters of paper, such as type of 

fiber, filler distribution, sheet formation, coating, and degree 

of surface sizing. Each of these parameters has a different 

degree of impact on the ink drying phenomena. Very important 

for ink jet printing is the surface tension of paper, which 

influences the wetting of the surface with liquid ink and 

affects two processes that occur simultaneously when an ink 

droplet hits the paper surface: spreading and/or penetration 

of the droplet. 16,17,26 The liquid wets (spreading) the surface 

of a solid substance if its surface tension L is lower than the 

surface tension S of the solid. Liquid surface tensions are 

directly measured, whereas solid surface tensions are commonly 

derived from contact angle measurements using 

semi-empirical equations, thus producing values that depend 

on the choice of contact angle test liquids and interpretative 

equations. 27,28 Surface and interfacial tensions are 

related to the contact angle by the Young equation, 26 

SV − SL = LV cos , 

where the subscript SV, SL, and LV refer to the solid/vapor 

and solid/liquid surfaces and the liquid/vapor interface, respectively. 

Surface tension is calculated on the basis of measuring 

the contact angles of two or three liquids of different 

polarities or surface charge. 17 When the surface tensions of 

two liquids are known and the contact angle measured with 

1 



Figure 5. Optical density D of 100% and 50% CMYK prints on papers1,2,and3. 

Measurements were made 7 days after printing. 

Figure 6. Color differences E * ab between the CMYK prints immediately 

after printing and 7 days of drying. 

these liquids, one can solve for the solid’s surface free energy 

components by writing an equation pair using either geometric 

or harmonic mean equations; note that the acceptable 

combinations of contact angles with a liquid pair are 

such that the increase in contact angles decreases the surface 

tension and polar component. 26–28 The surface tension was 

determined on the basis of the contact angle measurements 

for water and formamide. The total surface tension has been 

resolved into the dispersion part (van der Waals) and the 

polar part, using the geometric mean method equation: 26 

i 1 + cos i =2 i s disp + i s pol 1/2 

i =1,2, ..., 

where the first component, s disp , is due to the dispersion 

forces and the second, s pol , to the hydrogen bond and electrostatic 

forces. The subscript i is used to number the test 

liquids. The parameters and for test liquids are known 

and are given in Table II. During research, the contact angles 

of water W =72.8 mN/m and formamide 

2 

F =58.0 mN/m were measured after 2s, and the surface 

tension on the topside of papers was calculated. Figure 2 

represents the contact angle values after 2, 3, 4, or 5s. The 

calculated values of total surface tension and its disperse and 

polar part are presented in Table I. 26 Paper samples 1 and 2 

obtained high contact angle values, denoting that their surfaces 

are less absorbent for water. The quality wetting with 

the totally absorbent surface was already obtained with paper 

3 with more polar surface tension after 2s, in comparison 

with papers 1 and 2, which exhibited very low polar component 

of total surface tension. Sorption in those two papers is 

caused only by the dispersion component of the liquid surface 

tension. 

The Monitoring of Colorimetric Properties of Prints 

Colorimetric properties of a print represent an important 

criterion of print quality. 4,5,11,13,17 Since measured values 

change with time, standards for offset printing recommend 

spectrophotometric tests on prints dried 72 h after 

printing. 18,19 An appropriate standard for ink jet printing has 



Table II. Surface tension components for the test liquids water and formamide, mN/m. 

Liquid -Total -Dispersion part -Polar part 

Water 72.80 21.80 51.00 

Formamide 58.00 39.00 19.00 

Figure 7. Cross-section structure of black K prints on papers 1, 2, and 

3 magnification 160. 

not been developed yet. Monitoring of the color difference 

* 

E ab that appears on prints after a certain time due to ink 

drying proved the interdependence of ink drying speed and 

paper properties. In addition, differences among inks of different 

colors occur. Therefore, each paper was tested to 

evaluate the time necessary for the colorimetric properties of 

a print to stabilize. It was presumed that the appearance of a 

print does not change if the color difference amounts to 

* 

0.2 unit or less. The color difference E ab was calculated 

according to Eq. (3): 23 

E * = L * 2 + a * 2 + b * 2 1/2 , 

where L * =L * t−L * 0, a * =a * t−a * 0, and b * =b * t 

−b * 0 are the differences calculated for monitoring ink 

color of the print dried for time t t and the original (0) 

color print, where t=0. Individual color prints on paper 

samples were tested in order to determine the shortest period 

necessary for a color to dry. After brief monitoring (Fig. 

3 

3), the results proved that cyan (C) and yellow (Y) inks need 

more time to dry than other colors. 

Figures 3 and 4 show that color differences of the prints 

increase with time. We tried to estimate the time necessary 

for the drying process to be completed so the prints would 

stabilize and their color would remain thereafter unchanged. 

As shown, the results depend strongly of the paper and ink 

characteristics. The longer time of drying was found for the 

cyan (C) ink, which required at least 4 to 7 days, regardless 

of paper. Magenta (M) ink needed 7 days to stabilize on 

papers 1, 2, and 3; however, no major color differences were 

observed after 4 days. For yellow (Y) inks, it was found that 

the changes of color could be observed as long as 7 days on 

paper 3. Surprisingly, yellow ink seems to stabilize after only 

3 days on papers 1 and 2. As for black (K) prints, the results 

have shown that the samples with 100% coverage stabilize 

within one day as no evident color difference was observed 

later. This, however, was not the case with the 50% prints; 

they required more time to stabilize. This is obvious especially 

on paper 1. 

According to the results, the biggest changes in colorimetric 

properties of the prints occur in the first 60 min after 

printing. A suitable drying period of ink for ink jet printing 

can be estimated on the basis of the average time needed for 

the samples to stabilize and dry. The color of prints was on 

average changing for 4 or even 5 days. Therefore it can be 

assumed that the drying process has not been completed 

before 96 (or even 120) hours after printing. According to 

the ISO 2834 standard covering offset printing, samples 

should be dried in air and dry after 72 h. In comparison 

with this standard’s requirements, our results show that the 

drying process of an ink jet printing ink takes at least 1 day 

longer. In addition, the results of monitoring the drying process 

for a longer period show that the highest optical density 

was achieved on paper 3, but, at the same time, most of the 

color prints on paper 3 were stabilized only after 7 days. 

Therefore, we can assume that paper 3, absorbing the highest 

amount of ink, takes longer to dry. 

Colorimetric Properties of Dry Prints 

The results of the optical density measurements for 100% 

and 50% CMYK coloration are presented in Fig. 5. For comparison, 

the final color differences E ab after 7 days for 

* 

CMYK color prints with 50% and 100% coloration are 

shown in Fig. 6. Considerable deviations among samples 

were noted; we presumed that the optical density value, D,of 

prints should be at least 1.0 at 100% coloration and 0.5 at 

50% coloration. For 100% coloration, all colors except cyan 

exceeded the optical density value of 1.0. The lowest value 



was therefore obtained by cyan (C), slightly higher values by 

magenta (M) and yellow (Y) colors, and the highest by black 

(K). Comparision of optical density of prints and the observed 

color differences after 7 days shows that highest optical 

density is obtained for samples with the smallest color 

difference and vice versa. Main deviations were seen in the 

cyan print on paper 2. Paper 3 yields higher optical densities 

than the other papers. Such results were probably caused by 

the composition of the specialty coating on this paper that 

enables total bonding, absorption, and adsorption of the ink 

onto coating pigment particles. 

From Fig. 6 it can be seen that the extent of color differences 

that were observed on prints during the drying period 

depends on paper as well as ink characteristics and is 

connected with their structure and chemical composition, 

i.e., with the process of binding of certain colorant onto the 

paper surface. 

Cross Sections of Color Prints on Paper 

Figure 7 represents cross sections for black prints. It can be 

seen that the ink is adsorbed onto the surface coating of 

paper 3, whereas papers 1 and 2 have no such capability and 

therefore the ink is absorbed into the whole cross sectional 

structure of the papers. That causes a decrease in print quality 

expressed by lower values of optical density as well as 

bigger color differences after drying and confirms the results 

* 

of measured D and E ab shown in Figs. 5 and 6. 

CONCLUSIONS 

On the basis of comparative analysis of prints on selected 

papers it can be concluded that the sorptive properties of 

paper surface are of key importance for the quality of an ink 

jet print. Optimal values of sorption and water penetration 

have to be provided by the paper surface in order to achieve 

optimal bonding and adsorption of ink onto pigment particles 

or fibers. In addition, optimal surface tension of paper 

has to be achieved with regards to the surface tension of ink. 

On all paper samples, the most evident changes in their 

colorimetric properties occur during the first 60 min after 

printing. 

Prints are for the most part stabilized after 4 to 5 days, 

depending on color, i.e., composition of the ink. Due to 

ongoing changes in color, colorimetric measurements are 

recommended not sooner than 96 h after printing. 

Paper 3 with a special pigment coating produced the 

best results with respect to the quality of color prints and 

their optical density. Color differences between the prints on 

this paper were lower. The color changes of prints on this 

paper during the process of drying were considerably lower 

than the changes observed for the color prints on the other 

two papers, in spite of the fact that the time necessary for 

stabilization of the ink and to obtain the final color was 

found to be larger than for prints on papers 1 and 2. 


The research was performed as applied research project “Interaction 

between paper and colors in digital print technology” 

and financially supported by Ministry of Higher Education, 

Science and Technology and Slovene paper and 

printing industry. 

REFERENCES 

1 R. Knappich and A. M. Helbling, “Global markets: Competitive 

advantage through R&D?”, 21st PTS Coating Symposium R. Sangl, Ed. 

(PTS Symposium ST 301, Munich, Germany, 2003) pp. 4-1–4-13. 

2 M. Aikal, S. Nieminen, L. Poropudas, and A. Sesesto, “The end user 

aspects in print products development”, Proc. 30. IARIGAI Advances in 

Printing Science and Technology (FGA, Acta Graphica Publishers, Zagreb, 

2003) pp. 259–266. 

3 R. N. Jopson, “Coated inkjet papers and base stock effects—an 

overview”, PITA Coating Conference Proceedings (PITA, Zeebra 

Publishing, Manchester, 2003) pp. 125–133. 

4 H. P. Le, “Progress and trends in ink jet printing technology”, J. Imaging 

Sci. Technol. 42, 46 (1998). 

5 H. Ullrich, “The development of a pigmented multi-purpose ink jet 

paper”, PITA Coating Conference Proceedings (PITA, Zeebra Publishing, 

Mancester, 2003) pp. 135–138. 

6 G. Baudin and E. Rouset, “Ink jet printing: Effect of paper properties on 

print quality”, Proc. IS&T’s NIP 17 (IS&T Springfield, VA, 2001) pp. 

120–124. 

7 M. Klamann and M. Wedin, “Print quality and market potentional for 

ink-jet technology”, Proc.30. IARIGAI Advances in Printing Science and 

Technology (FGA, Acta Graphica Publishers, Zagreb, 2003) pp. 99–110. 

8 F. Eder, “Requirements for office communication papers of today and 

tomorrow”, Proceedings of the 29th International IARIGAI Research 

Conference (EMPA/UGRA, St. Gallen, 2002) pp. 17–26. 

9 K. Vikman, “Studies on fastness properties of ink jet prints on coated 

papers”, dissertation, Helsinki University of Technology, Espoo, 2004, p. 

95. 

10 C. Lie, W. Eriksen, and V. Matsegard, “Short-run printing—Influence of 

paper on print quality”, Proceedings of the 26th International IARIGAI 

Research Conference, Advances in Digital Printing (FOGRA/PTS, 

Munich, 1999) p. 2.2.3. 

11 W. Sobotka, “Digital printing—a comparison between electro 

photography and inkjet systems with regard to physical, chemical and 

printability limitations”, Proceedings of the 26th International IARIGAI 

Research Conference, Advances in Digital Printing (FOGRA/PTS, 

Munich, 1999) p. 4.7. 

12 W. Sobotka and N. Schuster, “New test methods for testing printability 

of ink-jet paper”, Proceedings of the 29th International IARIGAI Research 

Conference (EMPA/UGRA, St. Gallen, 2002) pp. 2.8. 

13 R. Sangl and J. Weigl, “On the interaction between substrate and 

printing ink for ink jet printing”, Proceedings of the 26th International 

IARIGAI Research Conference, Advances in Digital Printing (FOGRA/ 

PTS, Munich, 1999) p. 2.2.1. 

14 U. Lindquist and J. Heilmann, “The paper dependence of print quality in 

drop-on-demand ink jet printing”, Proceedings of the 26th International 

IARIGAI Research Conference, Advances in Digital Printing (FOGRA/ 

PTS, Munich, 1999) pp. 3.4. 

15 Y.-G. Tsai, M. Inoue, and T. Colasurdo, “The effect of sizing materials on 

the ink absorption in paper”, TAPPI 99 “Preparing for the next 

millennium” (TAPPI Press, Atlanta, GA, 1999) Book 1, pp. 111–122. 

16 M. Vaha-Nissi and J. Kuusipalo, Paper and Paperboard Converting, 

Chapter 3: Wetting and Adhesion in Paper and Board Converting (Fapet 

Oy, Helsinki, 1998) pp. 24–59. 

17 M. Von Bahr, J. Kizling, B. Zhmud, and F. Tiberg, “Spreading and 

penetration of aqueous solution and waterborne inks in contacts with 

paper and model substrates”, Proceedings of the 29th International 

IARIGAI Research Conference (EMPA/UGRA, St. Gallen, 2002) pp. 

87–102. 

18 O. Norberg, M. Andersson, and B. Kruse, “The influence of paper 

properties on colour reproduction and color management”, IS&T’s 

NIP19 (IS&T Springfield, VA, 2003) pp. 836–840. 

19 ISO 2834-2002. Graphic technology: Test print preparation for offset and 

letterpress inks (ISO, Geneva), www.iso.org. 

20 ISO 13655-1996. Graphic technology: Spectral measurement and 

colorimetric computation for graphic arts images (ISO, Geneva), 

www.iso.org. 

21 ISO 13656-2000. Graphic technology: Application of reflection 

densitometry and colorimetry to process control or evaluation of prints 

and proofs (ISO, Geneva), www.iso.org. 

22 G. N. Simonian and T. Johnson, “Investigation into the color variability 

& acceptability of digital printing”, Proceedings of the 28th International 



IARIGAI Research Conference, Advances in Color Reproduction (The 

Quebec Institute of Graphic Communication, Montréal, 2001) p. 4.6. 

23 B. Thompson, Printing Materials: Science and Technology (Pira 

International, Leatherhead, Surrey, 1999) pp. 410–431. 

24 G. Baudin, “Color control by densitometry approach with applications 

to offset and ink-jet printing”, Proceedings of the 28th International 

IARIGAI Research Conference, Advances in Color Reproduction (The 

Quebec Institute of Graphic Communication, Montréal, 2001) p. 3.4. 

25 O. J. Kallmes, “M/K Systems, Inc.—The Z-direction, non-uniformity of 

paper”, APR Europe 1, 40 (1991). 

26 R. Seppänen, M. Von Bahr, F. Tiberg, and B. Zhmud, “Surface energy 

characterization of AKD-sized papers”, J. Pulp Pap. Sci. 30, 70 (2004). 

27 R. J. Good, “Contact angle, wetting and adhesion: A critical review”, in 

Contact Angle, Wetability and Adhesion, K. L. Mittal, Ed., ISBN 90-6764- 

157-X (VSP, Utrecht, Netherlands, 1993). 

28 B. Thompson, Printing Materials: Science and Technology (Pira 

International, Leatherhead Surrey, 1999) pp. 177–180. 




Development of a Multi-spectral Scanner using LED Array 

for Digital Color Proof 

Shoji Yamamoto, Norimichi Tsumura and Toshiya Nakaguchi 

Department of Information and Image Sciences, Chiba University, Yayoi-cho, Inage-ku, 

Chiba, 263-8522, Japan 

Yoichi Miyake 

Research Center for Frontier Medical Engineering, Chiba University, Yayoi-cho, Inage-ku, 

Chiba, 263-8522, Japan 

E-mail: yamasho@graduate.chiba-u.jp 

Abstract. The authors have developed a multi-spectral scanner for 

accurately printing proofs that employs an LED array coupled with a 

photodiode array to measure the reflectance spectra. The system is 

composed of an LED array with five different spectral radiant distributions 

and 2048 silicon photodiodes with a Selfoc lens array (SLA) 

for imaging. Five types of LED were selected from among 40 types 

of commercially available LED with different spectral radiant distributions 

in order to minimize the average color difference E 94 be- 

* 

tween the measured and estimated reflectance spectra of 81 typical 

color charts. The multiple regression method based on the clustering 

and polynomial regression algorithm was introduced for highly 

accurate estimation of the spectral reflectance for printing. The results 

indicate that the average and maximum color differences E 94 

* 

between the measured and estimated reflectance spectra of 928 

color charts were 1.02 and 2.84, respectively. The scanner can 

measure the reflectance of prints having a 0.5 mm pitch resolution 

and a scanning speed of 100 mm/s. The field programmable gate 

array (FPGA) and digital signal processor (DSP) were introduced in 

order to accelerate the calculation of sensor calibration and the estimation 

of the reflectance spectra of the printed proof for practical 

and commercial use. As a result, the developed scanner could measure 

the reflectance spectra of the printed proof within 20 s. 



INTRODUCTION 

Color proofing has been widely used to evaluate and consider 

the color reproduction in printing, in order to provide 

a guarantee to customers regarding the quality of print based 

on the colorimetric color reproduction. In recent years, the 

availability of accurate digital color proofs via computer networks 

has reduced the cost and time associated with 

transportation. 1,2 

A color densitometry scanner is usually used to measure 

and digitize the color information of the color proof into R, 

G, B densities. 3–5 Printing proofs based on densitometric 

measurement are influenced by the illuminantion condition. 

For colorimetric color reproduction in the printing industry, 

it is necessary to compare color proofs and prints under the 

Received Jul. 3, 2005; accepted for publication Oct. 2, 2006. 

1062-3701/2007/511/61/9/$20.00. 

illuminant D50. 6,7 In the process of gaining approval by the 

customer, however, the use of D50 is not always practical. 

Recently, multi-spectral imaging 8–15 has been developed 

for accurate color reproduction under different illuminants. 

The reflectance spectra of the object are acquired in this 

imaging system for calculating the colorimetric values under 

arbitrary illuminants. Multi-spectral imaging is usually performed 

using five or more color filters for multi-band imaging. 

Typically, rotating filters are mounted in front of a 

monochrome CCD camera. 8–11 However, a great deal of 

time is required to rotate the filters with a mechanical wheel. 

Therefore, instead of rotating filters, a liquid crystal tunable 

filter (LFTF) may be used in multi-spectral imaging. 12–14 

This is appropriate for high speed measurement because the 

LFTF can change the spectral distribution of the filter, such 

as the peak wavelength and bandwidth, within several milliseconds. 

As a recently developed method for high-speed 

measurement, the CRISTATEL project 15 uses a small cask 

with filters and a linear CCD array detector, which provides 

10 ms scanning for each filter. However, these methods require 

a distance of more than 30 cm between the device and 

the object, which is not practical for factory use. In addition, 

it is necessary to satisfy the specifications of accuracy, compactness, 

and high speed measurement for creating digital 

color proofs in the print industry. 

We developed a multi-spectral scanner using an LED 

array and a photodiode array in order to accurately measure 

the spectral characteristics of the printing proof. A compact 

scanner can be achieved using LED illumination and an optical 

element, such as the Selfoc lens array. Conventional 

color filters are not necessary in this scanner because the 

LED emits light that has a band-limited spectral radiant distribution. 

Since the LED response time is very fast, highspeed 

measurement is possible by the timesharing control of 

each LED emission. 

In designing the multi-spectral scanner with an LED 

array, it is important to decide the number of LEDs and the 

spectral radiant distribution of each LED. The algorithm by 

which to decide the optimal combination of LEDs is explained 

in the third section. We develop the multi-spectral 

61

Yamamoto et al.: Development of a multi-spectral scanner using LED array for digital color proof 

scanner using the obtained optimal combination of LEDs 

and evaluate the accuracy of estimated reflectance spectra in 

the fourth and fifth sections, respectively. In order to improve 

the accuracy of the estimation, we also introduce additional 

algorithms using the clustering method and the 

polynomial regression method in the sixth section. Finally, 

concluding remarks are presented in the seventh section. 

COMPACT MULTI-SPECTRAL SCANNER USING AN 

LED ARRAY 

Figure 1 shows a schematic design of the proposed multispectral 

scanner. In order to satisfy the geometric conditions 

defined by the ISO or DIN standard 6 for the 0–45° method, 

the LED array is attached to a mount in order to illuminate 

the print from 45°, and the detector array is set to detect the 

light at 0° from the print. The Selfoc lens array (SLA) is 

inserted between the print and the detector in order to 

achieve a compact structure. 

In this system, a multiple-color type LED is used for 

multi-spectral imaging. Each emission for color can be controlled 

independently in this multiple-color LED. The analog 

responses of the photodetector for each color emission in 

the LED are converted to digital values, and the calibrated 

value P i x,y at position x,y illuminated by the ith LED is 

expressed as 

780 

SL i y,Rx,y, d − Dy 

1 

P i x,y =380 

, 1 

780 

Wr i 

SL i y,Wy, d − Dy 

380 

Figure 1. Schematic illustration of the multi-spectral scanner using an LED array. 

spectral radiant distribution of the ith LED at position y. 

The spectral reflectance Wy, is measured on the reference 

white plate at position y, and Dy is the measured 

response of the photodiode when all of the LEDs are 

switched off. The coefficient Wr i is used to compensate the 

difference between the reference white plate and the standard 

white corresponding to the ith LED. The practical use of the 

LED for color measurement requires two compensations, 

one for amplitude fluctuation and one for wavelength fluctuation. 

Equation (1) indicates a compensation for the amplitude 

fluctuation of the LED. A compensation for the 

wavelength fluctuation is taken into account in the LED selection 

in the third section. Equation (1) is applied for a 

large number of photodiodes in the multi-spectral scanner. 

In our system, we use a field programmable gate array 

(FPGA) for calculation because the FPGA has can perform a 

large number of simple, high-speed calculations. 

As mentioned above, each color emission in the LED is 

controlled by the timesharing process, and the responses of 

the photodetector for color emissions are ordered and 

where S is the spectral sensitivity of the photodiode, R 

is the spectral reflectance of the print, and L i y, is the 

Figure 2. Spectral radiant distribution of commercially available LEDs. 



streamed in the time series. The stream of responses is 

stored in memory for each set of color emissions in the 

pixel. Based on the stored set of color emissions, the spectral 

reflectance is estimated in the digital signal processor (DSP). 

The DSP is superior for calculating vector-matrix operation 

at high speed, i.e., for handling the stored responses in the 

memory. In the present paper, the multiple regression 

method is used for spectral estimation. 9 The estimation process 

for the multiple regression method is expressed simply 

as follows: 

Rˆ 380 

Rˆ 390 

] 

380,1 A 380,2 ¯ A 380,i 1 

A 390,1 

P 2 

] ] 

Rˆ A 

780=A 780,1 ¯ A 780,iP 

] 

P i, 

where Rˆ is the estimated reflectance at wavelength , and 

A ,i are the elements of the estimation matrix, which is determined 

from the relationship between the scanner response 

and the spectral reflectance of the samples. The 

sample should be chosen so as to represent the target prints 

and should be measured a priori. 

2 

Figure 3. Example of the shift of peak wavelength generated in the epitaxial 

deposition manufacturing process. 

Figure 4. Reflectance spectra of 81 color samples printed on coated 

paper. 

SELECTION OF LEDs 

In developing a multi-spectral scanner using an LED array, it 

is important to decide the number of LEDs and the spectral 

radiant distribution of the LEDs. In conventional multispectral 

imaging using the color filters, it is possible to optimize 

the spectral distribution of the filters 9 and produce 

the optimized filters in industry. However, the spectral radiant 

distribution of the LED has already been decided by the 

epitaxy process of the LED. Thus, it is not practical to optimize 

the spectral radiant distribution of the LED when designing 

the imaging system. In the present paper, we selected 

an LED combination from 40 types of commercially available 

LEDs in order to minimize the error between the original 

reflectance and the estimated reflectance. Figure 2 shows 

the spectral radiant distributions of the LEDs, which are 

normalized by the peak power and are obtained from the 

specifications of the LED. However, the peak wavelength of 

each LED is usually shifted by the fluctuations in the epitaxial 

deposition process during manufacture. Figure 3 shows 

typical examples of this fluctuation. This fluctuation must be 

taken into account in order to select the LEDs for robust 

design in the imaging system. 

In the following, we will explain the flow of the LED 

selection for the optimized robust imaging system. In the 

first step, the number of LEDs to be selected is i, and the 

flow is repeated while varying i from3to7,inorderto 

decide the optimal number of LEDs. Here, n is the combination 

of 40 items taken i at a time, and the evaluation 

process is repeated n times by changing the combination of 

LEDs. 

Next, the responses for the reflectance sample illuminated 

by the LED combination are obtained by computer 

simulation of the imaging system, and the calibrated responses 

are obtained by Eq. (1). The spectral reflectance is 

estimated from the calibrated responses by using the multiple 

regression method, as given by Eq. (2). The noise generated 

by the photo detector is usually added in the optimization 

process of the multiple regression method. However, 

this noise is ignored in our selection of the optimal LED 

because our system has a circuit for the compensation of the 

signal-to-noise ratio, as shown below in the fourth section. 

This circuit can provide an adequate signal-to-noise ratio by 

controlling the radiation time of the LED, even if a narrowband 

LED is selected. 

If the color difference between the measured reflectance 

and the estimated reflectance is greater than the permissible 

threshold, then the calculation progresses to the next evaluation 

process by changing the combination of LEDs. In addition, 

if the color difference is equal to or less than a permissible 

threshold, the color difference value and the 

combination of LEDs are recorded. 

The estimated reflectance is evaluated in comparison 

with the original reflectance of the sample. In the present 

paper, 81 samples for reflectance are examined for each 

evaluation of the LED combinations. These samples are halftone 

printed samples and white paper. The halftone samples 

have dot areas ranging from 10% to 100% in 10% pitches of 

C, M, Y, K, MY, CY, CM, and CMY, respectively. Figure 4 

shows the reflectance spectra of the 81 color samples, which 

are printed on coated paper and measured by a portable 

spectrophotometer (Gretag-Machbeth Spectro-Eye). The 

choice of these samples is related to the application of the 

digital color proof for offset print. 



Figure 5. Results of simulation for various numbers of LEDs D50 

illuminant. 

The criteria for evaluation should be set so as to meet 

the criteria used in practical applications for various types of 

papers and illuminants. For the criteria in the present paper, 

we first optimized the LED selection by using illuminant 

D50 and coated paper, which is the most popular combination 

in the graphics industry and is defined in ISO 13655. 6 

The criteria for other paper and illuminant combinations 

were evaluated using the optimal combination of LEDs, 

which was selected using illuminant D50 and coated paper. 

It is also necessary at evaluation to consider the fluctuation 

of the peak wavelength for the criteria. When the color 

* 

difference E 94 in the CIE L * a * b * color space 16 is equal to or 

less than a permissible threshold in the evaluation process, 

we add a ±10 nm variation to the peak wavelength for each 

LED in the process of estimation. The degree of variation of 

±10 nm is decided with a sufficient range from the measurements 

of LEDs, as shown in Fig. 3. Since the variations of 

−10 nm, 0nm, and +10 nm should be applied to each LED, 

we have 3 i variations for i LEDs. The maximum color difference 

E 94 indicates the maximum value obtained by the 

* 

calculated results of 3 i types of variation, which has −10 nm, 

0nm, and +10 nm fluctuations of peak wavelength at each 

LED, respectively. Finally, the maximum color difference 

* 

E 94 is used for the final evaluation value of the current 

combination of LEDs. 

Figure 5 shows the results of calculation for the variation 

of the number of LEDs. In this figure, the triangles 

* 

indicate the result of the maximum E 94 without the 

±10 nm fluctuation of the peak wavelength, and the squares 

indicate the results of the maximum E * 94 , which was used 

for robust assessment of the ±10 nm fluctuation. Both results 

indicate that the accuracy of estimation is improved as 

the number of LEDs increases. Five LEDs are necessary in 

order to estimate a spectral reflectance below the maximum 

E * 94 =2, which is calculated between the original reflectance 

spectra and the estimated reflectance spectra. Figure 6 shows 

the spectral radiant distributions of the best combination for 

three, four, five and six LEDs. The best number of LEDs was 

determined to be 5, and the peak wavelengths of the LEDs 

were obtained as 450, 470, 530, 570, and 610 nm, respectively. 

As mentioned above, five LEDs were determined to be 

effective for various printed papers and illuminant conditions, 

although the best selection was obtained by evaluation 

using only illuminant D50. We next evaluate the effectiveness 

of the estimation in detail. Table I shows the printed 

paper and illuminant conditions that are used to verify the 

influences of the printed paper and illuminant conditions on 

the estimation. The accuracy of the estimation is examined 

using “art paper” and “matte paper,” which are often used in 

the print industry. The estimation matrix is calculated using 

81 color samples of “coated paper” prints under the illuminant 

D50, as mentioned above, and the spectral reflectance 

is estimated from the response of the multi-spectral scanner 

for art and matte color samples. The accuracy of the estimation 

is also examined under A, C, D50, and D65. Figure 7 

* 

and Table II show the results as maximum E 94 between the 

Figure 6. Spectral radiance distribution of LEDs obtained by simulation using an LED array. 



Table I. Calculation conditions verified by the influence of the printed paper and the 

illuminant an LED array. 

LED No. Printed paper Illuminant 

Condition 1 

5 Coat D50 

5 Art D50 

5 Matte D50 

Condition 2 

5 Coat A 

5 Coat C 

5 Coat D50 

5 Coat D65 

* 

Table II. Results for maximum E 94 between the original reflectance and the estimated 

reflectance for each paper and illuminant. 

Printed 

paper 

* 

Condition Color difference E 94 

Illuminant 

Without 

fluctuation 

With 

fluctuation 

Coat D50 0.48 1.52 

Art D50 0.44 1.51 

Matte D50 0.28 0.89 

Coat A 0.52 1.55 

Coat C 0.50 1.53 

Coat D50 0.49 1.52 

Coat D65 0.44 1.52 

original reflectance and estimated reflectance in 81 color 

samples for each paper and illuminant. Note that the estimated 

reflectance for this result is calculated using the best 

selection among 450, 470, 530, 570, and 610 nm LEDs, 

which are obtained with illuminant D50 and coated paper. 

The black bar in this graph shows the results for maximum 

E 94 without ±10 nm fluctuation of peak wavelength, 

* 

* 

and the gray bar shows the results for maximum E 94 with 

±10 nm fluctuation of the peak wavelength. In Fig. 7(a), the 

estimation accuracies for art and matte papers are higher 

than that for coated paper, even if the estimation matrix was 

designed for coated paper. In general, the accuracy of estimation 

depends on the spectral gamut range, which is the 

low-dimensional linear space calculated by principal component 

analysis with training samples. Since the spectral gamut 

ranges of colors on art and matte papers are included within 

the spectral gamut range of colors on coated paper because 

of ink/media interactions, the estimated results of color on 

art or matte paper can be more accurately approximated. 

* 

Figure 7 and Table II show the results for maximum E 94 

between the original reflectance and the estimated reflectance 

in 81 color samples for each paper and illuminant. 

Note that the estimated reflectance is calculated using the 

best selection of 450, 470, 530, 570, and 610 nm LEDs, 

which was obtained using illuminant D50 and coated paper. 

We obtained highly accurate reproduction for illuminants A, 

C, and D65. Based on Figs. 7(a) and 7(b), it is confirmed 

empirically that the estimation matrix obtained for coated 

paper and illuminant D50 is also effective for other types of 

paper and illuminant. 

As a result, we found that the best number of LEDs is 

five, and the peak wavelengths of LEDs were obtained as 

450, 470, 530, 570, and 610 nm, respectively. 

Figure 7. Results of E * 94 for each type of paper and illuminant. Figure 8. Prototype multi-spectral scanner. 



Figure 9. Resulting images measured by the multi-spectral scanner: a 

color proof printed by press, b scanned image by a 450 nm LED, c 

scanned image by a 470 nm LED d scanned image by a 530 nm LED 

e scanned image by a 570 nm LED f scanned image by a 610 nm 

LED. Available in color as Supplemental Material on the IS&T website, 

www.imaging.org. 

DEVELOPMENT OF A MULTI-SPECTRAL SCANNER 

We developed a multi-spectral scanner using the best combination 

of five LEDs. Figure 8 shows a prototype multispectral 

scanner that can measure a 1024 mm800 mm 

print sample, and Fig. 9 (Available in color as Supplemental 

Material on the IS&T website, www.imaging.org) shows the 

resulting images that were measured by this scanner. The 

scanner consists of a sensor head with a detector, LED illuminations, 

and a processing circuit. The detector has a 2048 

photodiode array and an SLA is inserted between the print 

and the detector. A surface-mount-type LED is used in the 

scanner in order to make it more compact for practical use. 

Thirty-two sets of the five selected LEDs were used as 

the multi-band illumination. The peak wavelength of all 

LEDs was measured before mounting, and we used only 

LEDs that have fluctuations within ±10 nm of the peak 

Figure 10. Timing chart of the timesharing process for each LED. 

Figure 11. Block diagram and picture of the processing circuit. 

wavelength, corresponding to the designed center wavelength. 

These sets of three LEDs are aligned to illuminate the 

print from an angle of +45°. We were unable to align five 

types of LED as one linear array because the power of each 

LED was insufficient for sparse alignment for each type of 

LED. Therefore, the remaining LEDs were aligned to illuminate 

the print from an angle of −45°. 

In the system, the print is scanned twice: forward and 

backward. Three types of LED from +45° angles are used for 

illumination in the forward scan, and two types of LED from 

−45° angles are used for illumination in the backward scan. 

Each LED emission is controlled by the timesharing process 

to illuminate the print by one type of LED at each time. 

Figure 10 shows the timing chart of the timesharing process. 

For effective output level setting, the time of the measurement 

at each line is divided by the ratio of the LED power in 

order to determine the duration time for each LED. 

The scanner is capable of sampling 20481600 pixels 

to measure an image of 1024 mm800 mm with a pitch of 

0.5 mm. A standard white plate is mounted at the home 

position of the scanner, and the responses of darkness and 

whiteness are initially measured at the home position. The 

amplitude fluctuation of each LED is compensated using 

this initial measurement, as shown in Eq. (1). The analog 

response of the photodetector for each LED illumination is 

converted to 16 bits by an A/D converter, and the number of 

digital data reaches approximately 2048160052, 

corresponding to 33 mb, which is acquired by all of the 

pixels for each LED. The digital data are sent from the multispectral 

scanner to the processing circuit by a high-speed 

transmitter. The processing circuit performs the calculations 

of the calibration and multiple regression method as given 

by Eqs. (1) and (2). 



Figure 11 shows the processing circuit, which is composed 

of the FPGA, the memory, and the DSP. The calibration 

of amplitude fluctuation by Eq. (1) is performed at the 

FPGA in the time series, and the stream of responses is 

instantly stored in the memory. The calculation of the multiple 

regression method by Eq. (2) is performed at the DSP, 

which is superior for handling responses stored in memory. 

In this calculation, expressed in Eq. (2), we adopt distributed 

computation using six DSPs, where 342 pixels are assigned 

to each of the six DSPs. 

The scanning speed is designed to require 4000 s per 

0.5 mm pitch based on the architecture of the hardware in 

the developed multi-spectral scanner. The total number of 

scans required to measure a proof with a width of 800 mm 

and a pitch of 0.5 mm is 1600. In this system, approximately 

16 s is required for the multi-spectral measurement because 

the color proof is scanned forward and backward. Therefore, 

the total measurement time, including calculation and display 

for practical examination, is less than 20 s. 

EVALUATION OF THE DEVELOPED SYSTEM 

In this section, we evaluate and discuss the performance of 

the newly developed multi-spectral scanner. The multiple 

regression matrix for estimation is determined from the 81 

color samples on coated paper, and the spectral reflectances 

are estimated from the responses for 928 colors in the 

ISO12642 IT8/3 chart. Figures 12(a) and 12(b) show the 

examples of estimated reflectance spectra compared to the 

original reflectance spectra. The best estimation, shown in 

Fig. 12(a), achieves an acceptable accuracy over the entire 

wavelength. In contrast, the worst estimation, shown in Fig. 

12(b), fails to fit the spectral reflectance in the region, except 

for the center wavelength of the selected LEDs. In this case, 

five LEDs are insufficient to represent the spectral pattern. 

Figure 12(c) shows the color difference between the 

original and estimated reflectance spectra of 928 colors 

charts using the developed multi-spectral scanner. The average 

color difference E 94 is 1.23, and the maximum color 

* 

* 

difference E 94 is 4.07. In general, in the printing industry, 

the empirically acceptable average color difference is approximately 

2.5, and the maximum color difference is approximately 

3.0 in the CIE L * a * b * color space. 17,18 Therefore, 

the multi-spectral scanner developed using LEDs is considered 

to have sufficient accuracy with respect to average color 

difference, even though the maximum color difference exceeds 

the value of E * 94 =3.0. In the next section, we will 

improve the estimation method in order to reduce the maximum 

color difference. 

CLUSTERING AND POLYNOMIAL REGRESSION 

In this section, the clustering method and polynomial regression 

method 19 are applied to improve the accuracy of 

estimation with respect to the maximum color difference. 

Figure 13(a) shows the CIE a * b * diagram of estimated color 

for 928 samples, which are printed on coated paper and 

observed under illuminant D50. The estimated color is obtained 

using the multiple regression method. The triangles 

indicate the color samples having a color difference greater 

Figure 12. Results of estimation accuracy using the multiple regression 

method. a Best examples of spectral reflectance by estimation. b 

Worst examples of spectral reflectance by estimation ↓ center wavelength 

of LEDs used herein. c Color difference of 928 colors in the IT/8 

chart between the original and estimated spectral reflectance. d Histogram 

of the color difference of 928 colors in the IT/8 chart 

than E * 94 =2.5, and the dots indicate the other color 

samples. In this diagram, the triangles appeared in the red 

and green hue regions. Therefore, we apply the clustering 

method in these regions, which, in this paper, is performed 

simply by dividing the area by hue angle. For the results that 

exceed 2.5 with respect to color difference E * 94 , the centers 

of the hue angles for the green and red regions were calculated 

by the k-mean method. In each region, the upper limit 

of the hue angle is decided by adding the quantity obtained 

by multiplying the standard deviation of these hue angles by 

3 to the center of the hue angle, and the lower limit of the 

hue angle is decided by subtracting the quantity obtained by 

multiplying the standard deviation by 3 from the center of 

the hue angle. 



The polynomial regression method is expected to improve 

the accuracy of estimation because the error of estimated 

reflectance is caused by a nonlinear characteristic, 

which is not expressed by the multiple regression method in 

Fig. 12(b). This polynomial regression method is performed 

in order to add the squared response P i 2 in the calculation of 

the multiple regression matrix, as shown in Eq. (3), 

Rˆ 380 

Rˆ 390 

] 

2 

2 

380,1 A 380,2 ¯ A 380,i A 380,1 ¯ A 380,i 

A 390,1 

] ] 

Rˆ A 

780=A 780,1 ¯ A 780,i 

2 

 

1 

P 2 

] 

P P i 

2 

P 1 

] 

P 

2. 

i 

3 

Figure 13. Improved results of estimation accuracy using the clustering 

method. a Results for the estimated spectra in the CIE a * b * spaces. b 

Color difference of 928 colors in the IT/8 chart. c Histogram of the 

color difference of 928 colors in the IT/8 chart. 

For the clustering method, we first performed a preliminary 

estimation using 81 samples for training and 928 

samples for testing. As shown in Fig. 13(a), based on the 

obtained results, for the red sample within the angles of −5° 

to 40° in hue, the angle was classified from the 81 samples, 

and the green sample within the angles of 145° to 175° was 

also classified from the 81 samples. As a result, three estimation 

matrixes were constructed using the clustered red 

sample, the clustered green sample, and the 81 samples. In 

practical scanning, the first estimation of 928 samples for 

testing is executed using a color matrix of all 81 samples, and 

L * a * b * of the estimated spectral reflectance is calculated. For 

L * a * b * of the samples included in the red or green cluster 

area, the estimation is executed again using the corresponding 

estimation matrix. Figure 13(b) shows the color difference 

of 928 color samples based on this clustering method. 

The average color difference is improved to be E * 94 =1.04, 

and the maximum color difference is E * 94 =3.89. From 

these results, the clustering method is considered to be effective 

for improving the accuracy of estimation for spectral 

reflectance. However, we could not effectively reduce the 

maximum color difference using the clustering method. 

For the polynomial regression method, we use an estimation 

matrix calculated by Eq. (3) with 81 colors, and 

evaluation is performed using 928 samples. Figure 14(a) 

shows a comparison of the estimated reflectance spectra calculated 

by the polynomial regression method and by the 

multiple regression method. This example is for the same 

sample as in Fig. 12(b), which is the worst sample using the 

multiple regression method. The spectral pattern better fits 

the original reflectance than that obtained by the multiple 

regression method, and the color difference between the estimated 

reflectance and the original reflectance is improved 

using the polynomial regression method. 

Figure 14(b) shows the color difference of 928 color 

samples using the polynomial regression method. The results 

show that the average and maximum color differences 

are improved to E * 94 =1.02 and E * 94 =2.84, respectively. 

CONCLUSION 

We have developed a multi-spectral scanner using an LED 

array to construct an accurate digital color proofing system. 

For the system design, a robust technique was proposed to 

select LEDs from combinations of 40 commercially available 

* 

LEDs in order to minimize the color difference E 94 between 

the measured reflectance and the estimated reflectance. 

In this selection of LEDs, the fluctuation caused by 

the epitaxial deposition process during manufacture was 

taken into account. As a result of LED selection, we found 

that five LEDs are required in order to estimate spectral 

reflectance with E * 94 =2. The peak wavelengths of the LEDs 

were selected as 450, 470, 530, 570, and 610 nm and were 

independent of changes in the illuminant conditions. 

For practical verification in the printing industry, we 

constructed a prototype multi-spectral scanner using the 



industry because the average color difference was then 

E * * 

94 =1.02 and the maximum color difference was E 94 

=2.84. In the present study, we believe that this multispectral 

scanner system is very significant for obtaining accurate 

digital color proofs. 


This study was supported in part by SCOPE (Strategic Information 

and Communications R&D Promotion Programme) 

of the Ministry of Public Management, Home Affairs, 

Posts and Telecommunications of Japan. 

Figure 14. Improved results for estimation accuracy using the polynomial 

regression method. a Example of spectral reflectance estimation. b 

Color difference of 928 colors in the IT/8 chart. c Histogram of the 

color difference of 928 colors in the IT/8 chart. 

LED array. In the sensor head, the photodiode array, which 

has 2048 pixels, was used as a detector, and a Selfoc lens 

array was inserted for imaging between the object and the 

detector. In the processing circuit, the FPGA and the DSP 

were used to accelerate the calculation of sensor calibration 

and spectral reflectance estimation. The scanner has a pitch 

resolution of 0.5 mm and a scanning speed of 100 mm/s.In 

practical evaluation, we found that the measurement was 

completed within 20 s, including calculation and display. 

The spectral reflectance of the 928 color chart is used to 

evaluate the accuracy of the measurement and the estimation. 

The estimation procedure was determined by measuring 

the spectral reflectance of 81 typical color samples. Using 

the multiple regression method, we found the average color 

difference was E * 94 =1.23 and the maximum color difference 

was E * 94 =4.07. The clustering method and the polynomial 

regression method were also introduced in order to 

improve the accuracy of the estimated reflectance spectra 

compared with the multiple regression method. Among 

these methods, the polynomial regression method was found 

to be most effective for practical application in the printing 

REFERENCES 

1 Printnet: http://www.printnet.com.au/ 

2 H. Yamane, “Next generation digital archive system based on super high 

definition imaging database”, Proc. Electronic Imaging and the Virtual 

Arts 2000 GIFU, 14, 1–7 (2000). 

3 S. Suzuki, T. Kusunoki, and M. Mori, “Color characteristic design for 

color scanners”, Appl. Opt. 29, 5187–5192 (1990). 

4 G. Sharma and H. J. Trussell, “Set theoretic estimation in color 

characterization”, J. Electron. Imaging 5, 479–489 (1996). 

5 F. Konig, “Reconstruction of natural spectra from color sensor using 

nonlinear estimation methods”, Proc. IS&T’s 50th Annual Conference 

(IS&T, Springfield, VA, 1997) pp. 454–458. 

6 ISO13655, “Graphic technology Spectral measurement and colorimetric 

computation for graphic arts images” (1996). 

7 International Color Consortium (ICC), “Recommendations for color 

measurement, White paper #3”, http://www.color.org/ 

ICC_white_paper3measurement.pdf 

8 P. D. Burns and R. S. Berns, “Analysis multispectral image capture”, 

Proc. IS&T/SID 4th Color Imaging Conference (IS&T, Springfield, VA, 

1996) pp. 19–22. 

9 H. Haneishi, T. Hasegawa, A. Hosoi, Y. Yokoyama, N. Tsumura, and Y. 

Miyake, “System design for accurately estimating the spectral reflectance 

of art paintings,” Appl. Opt. 39, 6621–6632 (2000). 

10 M. Yamaguchi, T. Teraji, K. Ohsawa, T. Uchiyama, H. Motomura, Y. 

Murakami, and N. Ohyama, “Color image reproduction based on the 

multispectral and multiprimary imaging: Experimental evaluation”, 

Proc. SPIE 4663, 15–26 (2002). 

11 S. Helling, E. Seidal, and W. Biehlig, “Algorithms for spectral color 

stimulus reconstruction with a seven-channel multispectral camera”, The 

Second European Conference on Color Graphics Imaging and Vision 

(IS&T, Springfield, VA, 2004) pp. 229–262. 

12 J. Y. Hardeberg, F. J. Schmitt, and H. Brettel, “Multispectral image 

capture using a tunable filter”, Proc. SPIE 3963, 77–88 (1999). 

13 F. H. Imai, M. R. Rosen, and R. S. Berns, “Comparison of spectrally 

narrow-band capture versus wide-band with priori sample analysis for 

spectral reflectance estimation”, Proc. IS&T/SID 8th Color Imaging 


14 J. Y. Hardeberg, F. Schmitt, and H. Brettel, “Multispectral color image 

capture using a liquid crystal tunable filter”, Opt. Eng. (Bellingham) 41, 

2532–2548 (2002). 

15 A. Ribes, H. Brettel, F. Schmitt, H. Liang, J. Cupitt, and D. Saunders, 

“Color and multispectral imaging with the CRISTATEL muitispectral 

system”, Proc. IS&T’s PICS (IS&T, Springfield, VA, 2003) p. 215. 

16 R. McDonald and K. J. Smith, “CIE94 a new colour-difference formula”, 

J. Soc. Dyers Colour. 111, 376–379 (1995). 

17 ISO/TC130 Activities and ISO Standards, “The Standardization of 

Graphic Technology, Japan Printing Machinery Association”, May, 2005; 

http://www.color.org.JapanColor2005English.pdf 

18 ASTM, “Method E 97-53 T” (1953). 

19 N. Tsumura, H. Haneishi, and Y. Miyake, “Estimation of spectral 

reflectances from multi-band images by multiple regression analysis”, 

Jpn. J. Opt. 27, 384–391 (1998) (in Japanese). 




Spectral Color Imaging System for Estimating Spectral 

Reflectance of Paint 

Vladimir Bochko † 

Department of Information Technology, Lappeenranta University of Technology, P.O. Box 20, 

53851 Lappeenranta, Finland 

E-mail: vbotchko@gmail.com 

Norimichi Tsumura and Yoichi Miyake 

Department of Information and Image Sciences, Chiba University, 1-33 Yayoi-cho, 

Inage-ku, Chiba 263-8522, Japan 

Abstract. In this paper, the analysis methods used for developing 

imaging systems estimating spectral reflectance are considered. 

The chosen system incorporates an estimation technique for spectral 

reflectance. Several traditional and machine learning estimation 

techniques are compared for this purpose. The accuracy of spectral 

estimation with this system and each estimation technique is evaluated 

and the system’s performance is presented. 



INTRODUCTION 

In this paper, the analysis methods used for developing imaging 

systems estimating spectral reflectance are considered. 

The estimation of spectral reflectance determines performance 

of a high quality color imaging system which is required 

for digital archives, network museums, e-commerce, 

and telemedicine. 1 Especially the design of a system for accurate 

digital archiving of fine art paintings has awakened 

increasing interest. In such a system the digital image is easily 

examined by using a broadband network. The visitors to 

museums, art experts and artists would be able to appreciate 

a variety of paintings at any viewing site regardless of where 

those paintings are located. In addition, archiving the current 

condition of a painting with high accuracy in digital 

form is important to preserve it for the future. Several research 

groups worldwide have been working on these 

problems. 2–14 

Conventional color imaging systems have some limitations, 

namely dependence of images on the illuminant and 

characteristics of the imaging system. The imaging systems 

based on spectral reflectance, unlike the conventional systems, 

are device independent and capable of reproducing the 

image of the scene under any illumination conditions. Also, 

these systems can incorporate the color appearance charac- 

† 

Current address: 51 Brigstock Road, Thornton Health, Surrey CR7 7JH, 

UK. 

Received Jun. 9, 2005; accepted for publication Jun. 7, 2006. 

1062-3701/2007/511/70/9/$20.00. 

teristics of the human visual system. Owing to the fact that 

spectral characteristics are smoothed, the high-dimensional 

spectral reflectance is accurately represented by a small number 

of channel images. 15–17 Therefore, the task of spectral 

estimation includes statistical analysis of the reflectance 

spectra and minimization of the estimation error. The 

choice of error measures is a general topic of broader interest, 

and choices are sometimes contrary in impact. In the 

archival realm, ramifications of optimizing on RMSE versus 

color difference may depend on applications. For example, 

spectral optimization may better enable the identification of 

colorants used while color difference optimization may yield 

superior visual reproductions. 

The traditional techniques used for the estimation involve 

matrix-vector computation and usually assume a linear 

model of the data. Although the approach based on linear 

algebra and a nonlinear data model is proposed in the 

literature, 4 machine learning techniques seem appealing. 

They estimate spectra of the scene, incorporate the data 

nonlinearity, and involve training and prediction procedures. 

Therefore, a neural networks-based method for spectral reconstruction 

has been proposed by Ribes et al. 18 The tested 

methods are superior to the pseudoinverse based estimation 

method with quantization noise. Without noise the traditional 

methods predict better than the neural network because 

of the highly linear relationship between spectral sets 

used for training and prediction. To provide color constancy 

a Bayesian approach of the estimation method is proposed 

by Brainard and Freeman. 19 Since the Bayesian approach is 

computationally demanding, the submanifold method for 

spectral reflectance estimation that is an intermediate solution 

between the Bayesian approach and linear estimation 

methods is described by DiCarlo and Wandell. 20 The 

method extends the linear methods and introduces an additional 

term incorporating the nonlinearity of the data. The 

method uses a piecewise linear way to represent the nonlinear 

data structure and reduces the error value 12% in comparison 

with a linear method. It is important that the 

method particularly reduces large linear errors. The limitation 

of the method is that it needs a large training set and is 

70

Bochko, Tsumura, and Miyake: Spectral color imaging system for estimating spectral reflectance of paint 

insufficient when the data structure is a one-to-many mapping. 

The properties of the methods considered in this paper 

are quite close to the submanifold approach 20 and one of the 

learning algorithms based on Wiener estimation also gives a 

piecewise linear solution. 

Recently, many advanced machine learning techniques 

using neural networks and support vector machines have 

been introduced and combined in the libraries that are convenient 

for the purpose. For example, for building the estimation 

methods using ready-made machine learning algorithms, 

one can obtain theoretically founded algorithms, a 

unified workflow for a current and future study, and a rich 

set of methods that provide flexibility for applicationoriented 

research. In this paper, the neural networks algorithms 

from the Netlab library 21,22 will be used. They include 

regression, clustering, and pattern recognition methods. 

Many of these methods are density models based on a likelihood 

that is important for recognition and convenient for 

comparison with other methods. 

In this study, we statistically analyze the reflectance 

spectra of color patch sets of oil and watercolor paintings 

without noise characteristics, develop three machinelearning 

based methods, and compare them with three traditional 

methods with a synthetic data set and the real color 

patch sets, as well. The traditional methods are linear estimators 

based on low-dimensional principal component 

analysis (PCA) approximation and Wiener estimation, and a 

nonlinear estimator based on multiple regression approximation. 

The machine learning methods extend the traditional 

methods for estimating a nonlinear data structure. 

They include two nonlinear methods based on nonlinear 

principal component analysis and regression analysis and the 

method using piecewise linear Wiener estimation. The 

method utilizing nonlinear PCA and the method exploiting 

piecewise linear Wiener estimation are novel methods. To 

develop an imaging system, two measures are used for estimation 

accuracy: spectral color difference (RMSE) and colorimetric 

color difference (CIE E 94 ). The former is better for 

archiving spectral reflectance and the latter is better for 

evaluating the appearance of the art paintings under a specific 

illumination to human observers. 

The paper is arranged as follows: In the following section, 

we formulate the generalized reconstruction of spectral 

reflectance from a multichannel image in imaging systems 

with a reduced number of channels. Next, we describe three 

traditional methods and three machine learning methods. 

Then we present the results of the statistical analysis of the 

reflectance spectra of the color patches. Later on, an experiment 

with synthetic data and reflectance spectra of the color 

patches is described. Finally, the experimental results are discussed 

and concluding remarks are presented. 

FORMULATION OF THE SPECTRAL 

REFLECTION ESTIMATION 

Figure 1 shows the image acquisition system. The system 

consists of a single chip, high quality charge coupled device 

(CCD) camera and a rotating color wheel comprising several 

Figure 1. The image acquisition system. 

color filters. The response at position x,y of the CCD 

camera with the ith color filter is expressed as follows: 3 

v i x,y = t i ESrx,y,d + n i x,y, 

i =1, ...,m, 

where t i , E, S, and rx,y, are the spectral transmittance 

of the ith filter, the spectral radiance of the illuminant, 

the spectral sensitivity of the camera, and the spectral 

reflectance of a painting, respectively; n i x,y denotes additive 

noise in the ith channel image, and m denotes the total 

number of channels. 

For mathematical convenience, each spectral characteristic 

with l wavelengths is expressed as a vector or a matrix. 

Using vector-matrix notation, we can express Eq. (1) as 

follows: 

vx,y = T T ESrx,y + nx,y, 

where T denotes a transposition, v is an m1 column vector 

representing the camera response, r is an l1 column 

vector representing the spectral reflectance of the painting, 

T=t 1 ,t 2 ,...,t m is an lm matrix in which each column t i 

represents the transmittance of the ith filter, and E, S are the 

ll matrices that correspond to the spectral radiance of the 

illuminant and the spectral sensitivity of the CCD camera, 

respectively. 

Further for the sake of simplicity, x,y from v, r, and n 

are omitted. Equation (2) is rewritten as an overall, linear 

system matrix F=T T ES with ml elements 

v = Fr + n. 

The response of the spectral CCD camera v without a 

noise term is as follows: 

v = Fr. 

We will call the space spanned by r a spectral space and 

the space spanned by v a sensor space or subspace. The 

estimation of reflectance spectra is obtained as follows: 

1 

2 

3 

4 



rˆ = Gv, 

where G is a matrix depending on the estimation method 

used. In the next sections, six estimation methods are 

considered. 

TRADITIONAL ESTIMATION TECHNIQUES 

Three approaches are usually used for spectral sensor design. 

The estimation techniques of reflectance spectra include: the 

method based on PCA (low-dimensional approximation) 

(PCE), the method based on Wiener estimation (WE), and 

the method using multiple regression approximation 

(MRE). 4 

The Method Based on PCA 

Using spectral reflectance of the training set r a covariance 

matrix is computed as follows: 

C = Er − Err − Er T , 

where E is an expectation operator. 

An eigendecomposition of the covariance matrix C determines 

the matrix B=b 1 ,b 2 ,...,b k , the columns of which 

are k eigenvectors corresponding to the first k largest eigenvalues. 

The spectral reflectance is approximated as follows: 

r Bw, 

where w is a vector of principle components (PCs), 

w=w 1 ,w 2 ,...,w k T and km. 

The spectral camera response given by Eq. (4) can be 

represented by another expression as follows: 17 

v = FBw. 

The PCs are determined as follows: 

w = FB −1 v. 

Using Eqs. (7) and (9) the estimation matrix G is as 

follows: 

G = BFB −1 . 

5 

6 

7 

8 

9 

10 

The estimate of the spectral reflectance of the painting 

is as follows: 

rˆ = Gv = BFB −1 v, 

11 

where the data is centered by v←v−EFr where ← means 

that the expression on the right is first calculated and then 

replaces the expression on the left. Finally, the mean value is 

added as follows: 

rˆ = rˆ + Er. 

12 

Better accuracy of estimation can be obtained with 

Wiener estimation, which is considered next. 

The Method Using Wiener Estimation 

The Wiener estimation method minimizes the overall average 

of the square error between the original and estimated 

spectral reflectance. 3 For this method, the correlation matrices 

R rr of painting spectra and noise R nn are first computed, 

and consequently, the estimation matrix is the following: 3 

G = R rr F T FR rr F T + R nn −1 . 

The estimate is as follows: 

rˆ = Gv = R rr F T FR rr F T + R nn −1 v. 

13 

14 

If noise is not considered, the estimation matrix is as 

follows: 3 G = R rr F T FR rr F T −1 . 15 

And the estimate is as follows: 

rˆ = Gv = R rr F T FR rr F T −1 v. 

16 

In this study, the Wiener estimation without consideration 

of noise is used. The Wiener estimation gives good 

accuracy for linear data. If the data is nonlinear, the technique 

based on multiple regression analysis is used. 

The Method Using Multiple Regression Analysis 

In the case of nonlinear data, multiple regression analysis 

gives better results than Wiener estimation. 4 

In the MRE method, the extended data matrix V of 

painting spectra is first defined through the data components 

and their extended set of higher-order terms as 

follows: 4 

V = v 1 , ...,v m ,v 1 v 1 ,v 1 v 2 ... ,higher-order terms, ... , 

where denotes element-wise multiplication. 

Then the estimation matrix is given as follows: 

G = RV T VV T −1 , 

17 

18 

where R is a matrix, the columns of which are presented by 

n spectral samples given by 

R = r 1 ,r 2 , ...,r n . 

19 

According to the literature, 4 the estimation matrix G used in 

MRE is equal to the noiseless variant of the Wiener estimation 

matrix. 

Finally 

rˆ = GV = RV T VV T −1 V. 

20 

Owing to the fact that new advanced machine learning 

algorithms are especially relevant for working with nonlinear 

structured data, the machine learning techniques are next 

discussed for spectral estimation. 

MACHINE LEARNING ESTIMATION TECHNIQUES 

By analogy with the traditional estimation methods, three 

machine learning techniques are proposed. They include the 

method based on regressive (nonlinear) PCA (RPCE), the 

method based on piecewise linear Wiener estimation 

(PLWE), and the method using regression analysis (RE). 

Equations. (1)–(5) are valid for all machine learning 

methods. 



The Method Based on Regressive PCA 

The spectral camera response is computed in the following 

way: 

v = FBfw, f , 

21 

where f is a nonlinear vector-valued mapping function 

and f is a parametric vector. 

Then, PCs are defined by the following equation: 

w = hFB −1 v, h , 

22 

where h is an inverse function, h =f −1 , h is a parametric 

vector, and v←v−EFr. 

The mapping function h and parametric vector h 

are computed using a machine learning algorithm for 

regression. 21 In consequence, the spectral estimate of the 

painting is as follows: 

rˆ = BhFB −1 v, h . 

Finally, the mean value is added as follows: 

rˆ ← rˆ + Er. 

23 

24 

In practice, this method involves a low-dimensional 

subspace and a higher-dimensional subspace including the 

low-dimensional subspace. For the low-dimensional subspace, 

where w k =w 1 ,w 2 ,...,w k T , the mapping is as follows: 

w k = hFB −1 v, h = FB −1 v, 

where v←Fr−EFr. 

For the higher-dimensional subspace, where 

w p = w k ,w k+1:p T = w 1 ,w 2 , ...,w k ,w k+1 , ...,w p T , 

25 

26 

the mapping is done for the higher-order (or weak) PCs as 

follows: 

w k+1:p = hFB −1 v, h = hw k ,. 

27 

Thus the method uses the low-order real PCs and the 

higher-order approximated PCs. 

The Method Using Piece-Wise Linear Wiener Estimation 

In this section, the other machine learning algorithm for 

piece-wise linear Wiener estimation is discussed. The main 

idea of the method is to separate the data structure into 

parts which are suitable for linear approximation and each 

part is then estimated by using the linear Wiener estimation 

method. 

For data separation, the clustering algorithm is first required. 

The data is divided into several clusters v i using the 

Gaussian mixture model 21 in a sensor space where i is an 

index of the cluster. Then for the data of each cluster Wiener 

estimation is utilized. Using the labels of the data it is easy to 

compute the cluster covariance matrix in the spectral domain 

needed for estimation. When the ith cluster covariance 

matrix C i of painting spectra is known, the spectral estimate 

for the ith cluster is as follows: 

rî = G i v i = C i F T FC i F T −1 v i , 

where v i ←v i −EFr i . 

Finally, the mean value is added as follows: 

rî ← rî + Er i . 

28 

29 

The estimation procedure is sequentially repeated for all 

clusters. 

The Method Using Regression Analysis 

The estimation method based on regression analysis is similar 

to the multiple regression approach. The difference is 

that nonlinear mapping is used instead of linear mapping 

and the higher-order terms are not synthesized. For regression 

analysis based on machine learning the estimate is given 

as follows: 

rˆ = gv,, 

30 

where g is a nonlinear vector-valued mapping function and 

is a vector of parameters. 

Then, an ith entry is defined as follows: 

rî = g i v,. 

31 

There are several regression algorithms 21 but only the 

regression method based on the radial basis function (RBF) 

is used in this study for the RE and RPCE methods. The 

reason is that the RBF method is relatively fast and performs 

well. 

ADDITIONAL TECHNIQUES 

All machine learning algorithms may need additional techniques 

to help in parameter adjustment. 

The regressive PCA method used in this study is a technique 

which combines the PCA and nonlinear regression 

methods. 23 In general, the methods utilized in both approaches 

to detect the underlying dimensionality of the data 

can be combined. For PCA, this is an analysis of the residual 

energy depending on a number of PCs. Furthermore, for 

regression methods this is automatic relevance determination 

(ARD). 21 The ARD method defines the statistical dependence 

between the PCs, and, in the case of the dependency 

between the tested components and a target 

component, the tested components are relevant to approximating 

the target component. However, this technique will 

not be used in this study. For the regressive PCA the number 

of real PCs will be given and a number of approximated PCs 

will be used as free parameters. 

The piecewise linear Wiener estimation approach needs 

to determine the number of linear components for use in a 

clustering procedure. This is done based on the model selection 

of the mixed distribution. 24 After that the Gaussian 

mixture model 21 with a given number of clusters is used to 

extract linear components. 



STATISTICAL PROPERTIES OF 

REFLECTANCE SPECTRA 

For statistical analysis of the spectral reflectance of paintings 

we use five sets of color patches of oil or watercolor paint as 

follows: set A, 336 patches of paint (reflectance of paint); set 

B, 60 patches of paint (Turner acryl gouache); set C, 60 

patches of paint (Turner golden acrylics); set D, 91 patches 

of paint (Kusakabe oil paint); and set E, 18 patches of paint 

(Kusakabe haiban). All sets were extracted from the standard 

object color spectral database constructed by the Spectral 

Characteristic Database Construction Working Group. 25 

These sets have a spectral range of 400–700 nm and 

samples are evenly taken at 10 nm. 

Set A is used for training the algorithms and Sets B–E 

are used for prediction of spectral reflectance. Therefore, 

linear and nonlinear principal component analysis was carried 

out only for set A. According to a previous publication, 3 

five PCs of linear PCA are good enough for accurate spectral 

estimation. Hence, spectral set A and its first five PCs that 

have a residual energy of 0.16% are analyzed and shown in 

Figs. 2 and 3, respectively. 

If regressive PCA is applied to utilize the five real PCs 

and several approximated PCs of set A, the average RMSE 

value of the spectral approximation is reduced (Fig. 4). This 

illustrates the fact that there is a way to improve the degree 

of accuracy for representing spectra by incorporating the 

nonlinearity of the data. 

EXPERIMENT 

Synthetic Data 

In this section, the nonlinear dataset is first synthesized and 

then all methods for spectral estimation are tested with a 

synthetic set. It is assumed that one channel response is used 

while the data simulating spectra is two dimensional. The 

purpose of the test is to show the feasibility of the method to 

work with data which has a nonlinear structure. 

Thus two data components are generated for the test. 

The first component x 1 is uniformly distributed in the range 

Figure 2. Reflectance spectra of set A paint patches. 

Figure 4. The average RMSE of spectral approximation for set A using 

regressive PCA. The first five components are given by PCA and the 

components 6–10 are approximated by regressive PCA. 

Figure 3. First five principal components of set A paint patches. 

Figure 5. The estimation results for the synthetic data and different estimation 

methods. 



−0.2–0.5 and another one is x 2i =x 1i −0.5. 4 Finally, a zeromean 

Gaussian noise with standard deviation 0.007 was 

added to the generated components. The estimation result of 

the synthetic data is presented in Fig. 5. A vector F, avector 

b 1 , that is a first PCA eigenvector from B, and the curve 

corresponding to an underlying subspace are shown in Fig. 

5. The original (synthesized) data and the estimates for each 

method are shown by the lines of dots in Fig. 5. 

Although the WE method is superior to the PCE based 

method, the PCE and WE methods give poor estimates for 

the data. The MRE, RPCE, and PLWE methods are relatively 

good for estimation. The RE method gives the best result 

from among these methods. 

Real Data 

An experiment was conducted with sets A–E described 

above. Set A is used for training while the other sets are used 

for prediction. The spectral transmittance characteristics of 

the separation filters used in a CCD camera are given in Fig. 

6. The spectral sensitivity of a CCD area sensor (Phase One 

3072horizontal pixels2060vertical pixels, 14 bits) is 

presented in Fig. 7. The illumination source is D65. 

The parameters used in the test are the following: The 

five PCs are exploited for PCE and RPCE. In addition, the 

RPCE approach uses the PCs approximating the real sixth, 

seventh, eighth, and ninth PCs. For the PLWE method a 

mixture of Gaussian components is used for clustering 

where the number of components is defined in a test based 

on the model selection of the mixed distribution. The MRE 

technique uses terms beginning with the first-order ones to 

the second-order ones. For the RE method, regression is 

based on the radial basis function using the Gaussian function; 

20 neurons and seven iterations are used in this case. 

A variational Bayesian model selection method for the 

mixture distribution 24 in the sensor space defines the number 

of components for the PLWE method. For this, the program 

is rerun ten times. The results are presented in Table I 

where the first row shows the test number and the second 

row shows the number of components determined by the 

algorithm. Figure 8 illustrates the variational likelihood 

bound over the model selection of 336 paint spectra (set A). 

Initially, the model has ten Gaussians. The vertical lines 

show the removal of the components from the model. Finally, 

two components are selected. 

If the estimation values of spectral reflectance are less 

than zero or greater than one then they are equalized to zero 

or one, respectively 

In Tables II and III, the average and maximum RMSE 

values for each set are given for the traditional methods and 

methods based on machine learning algorithms, respectively. 

In Tables IV and V, the average and maximum CIE 

E 94 values for each set are given for the traditional methods 

and methods based on machine learning algorithms. 

In general, the results presented in Tables II–V demonstrate 

that for the RMSE values the machine learning methods 

give slightly better results than their traditional opposite 

methods while the traditional methods have smaller CIE 

Table I. Number of components for piecewise linear Wiener estimation. 

Test number 1 2 3 4 5 6 7 8 9 10 

Number of components 2 2 1 1 2 2 2 2 2 2 

Figure 6. The spectral transmittance characteristics of the filters. 

Figure 7. The spectral sensitivity of the camera. 

Figure 8. The variational likelihood bound over the model selection of 

336 paint spectra set A. 



Table II. Average and maximum in parentheses RMSE values for PCE, WE, and MRE. 

PCE WE MRE 

Set A 0.0516 0.2458 0.0155 0.1633 0.0123 0.1159 

Set B 0.0836 0.3952 0.0346 0.1712 0.0324 0.1732 

Set C 0.0889 0.3469 0.0466 0.2478 0.0397 0.2158 

Set D 0.0917 0.4083 0.0403 0.2304 0.0352 0.2075 

Set E 0.0917 0.3136 0.0330 0.1416 0.0281 0.1199 

Table III. Average and maximum in parentheses RMSE values for RPCE, PLWE, and 

RE. 

RPCE PLWE RE 

Set A 0.0512 0.2447 0.0142 0.1522 0.0123 0.1047 

Set B 0.0834 0.3928 0.0343 0.1683 0.0315 0.1731 

Set C 0.0887 0.3452 0.0450 0.2350 0.0379 0.2010 

Set D 0.0912 0.4066 0.0376 0.2209 0.0349 0.1992 

Set E 0.0910 0.3122 0.0339 0.1185 0.0275 0.1062 

Table IV. Average and maximum in parentheses CIE E 94 values for PCE, WE, and 

MRE. 

E 94 values. The exception is the RE method which has 

better prediction in comparison with the other methods for 

the maximum error of the color difference. 

The methods are also tested with respect to computational 

time. The CPU time in seconds for set A is presented 

in Table VI. MATLAB 6.5, the Intel Pentium III Processor, 

1066 MHz and 248 MB of RAM are used in the test. For the 

various algorithms, the CPU time is given separately for 

training (upper row) and prediction (lower row). In Table 

VI, zero values are given for the CPU times which are very 

small (this corresponds to several matrix-vector multiplications). 

The test shows that the traditional methods are faster 

than the machine learning methods. However, the prediction 

time for the machine learning methods is relatively short. 

To see whether any nonlinearity is presented in the estimated 

spectra we measure the average RMSE value after 

estimation of spectral reflectance using PCA and RPCA. The 

results are shown in Table VII for PCA with the five PCs 

(upper number) and for RPCA with the five real PCs and 

five approximated (from 6 to 10) PCs (lower number). 

Then, the ratio between these two RMSE values is determined 

and presented in Table VIII. 

From Table VIII, one can see that the RE and RPCE 

methods have ratio values close to the original data set. The 

MRE and PLWE methods give results which are farther from 

the original data set. The PCE and WE ratio values are the 

most different from the original data in comparison with the 

other methods. 

From among the traditional methods the method based 

on MRE produces the best result. The method has small 

RMSE and CIE E 94 values in the training set and sets used 

for prediction. While the RMSE values for all machine learn- 

PCE WE MRE 

Set A 0.72 13.65 0.17 4.03 0.15 1.68 

Set B 2.96 21.00 0.58 2.84 0.54 2.13 

Set C 2.36 15.42 0.80 4.08 0.59 4.21 

Set D 2.43 19.24 0.71 5.18 0.55 3.37 

Set E 1.32 3.57 0.37 2.34 0.31 1.18 

Table V. Average and maximum in parentheses CIE E 94 values for RPCE, PLWE, 

and RE. 

RPCE PLWE RE 

Set A 0.81 14.89 0.16 3.46 0.17 3.16 

Set B 3.34 23.15 0.67 2.65 0.59 2.65 

Set C 2.51 14.90 1.033 8.47 0.82 3.47 

Set D 2.71 20.86 0.8623 8.19 0.74 2.92 

Set E 1.89 5.14 0.57 2.00 0.71 2.79 

Table VI. CPU time in seconds. 

PCE WE MRE RPCE PLWE RE 

0.04 0.0 0.01 0.35 0.38 6.49 

0.0 0.0 0.01 0.03 0.22 0.18 

Table VII. Average RMSE value after spectral estimation for PCA with five PCs upper 

number and for RPCA with five real components and five approximated components 

lower number. 

Set A PCE WE MRE RPCE PLWE RE 

0.00941 0.00441 0.00043 0.00728 0.00614 0.00453 0.00807 

0.00772 0.00422 0.00048 0.00539 0.00479 0.00423 0.00626 

Table VIII. Ratio between the RMSE values for PCA and RPCA. 

Set A PCE WE MRE RPCE PLWE RE 

1.21 1.04 0.88 1.35 1.28 1.07 1.29 



ing methods are slightly better in comparison with the traditional 

methods, the CIE E 94 values of the methods based 

on machine learning except the RE method are higher. The 

overall means of average color differences for the traditional 

methods are 1.95 (PCE), 0.52 (WE), and 0.42 (MRE) and 

for the learning methods 2.25 (RPCE), 0.65 (PLWE), and 0.6 

(RE). Thus, the color differences using the machine learning 

methods are smaller than the differences between the traditional 

methods. The RE method incorporates nonlinearity of 

data as is clearly seen from Table VIII. The generalization of 

the data given by the RE method is very good in comparison 

with the other methods. This follows from the maximum 

CIE E 94 values. However, given the processing and execution 

times the MRE method gives a better average, and in 

two out of five cases smaller maximum color difference errors 

than the RE method. Although the traditional methods 

are less time consuming than the machine learning methods, 

the prediction time for the learning methods is short 

enough. 

In general, the traditional methods look more desirable 

than the machine learning methods. This is contrary to the 

initial expectation from the result shown in Fig. 5 where the 

learning methods appear superior to the traditional methods. 

This can be explained as follows. In this study the sensor 

space (subspace) dimensionality is defined by the five 

given filters. Although the subspace is not optimal (close to 

optimal) its dimensionality is rather high. Recently, it was 

shown that for reflectance spectra the dimensionality of the 

nonlinear subspace is approximately three. 26 Thus, one can 

expect that for spectral imaging systems having the lowdimensional 

sensor space or fewer channels the learning 

based methods are more efficient. We will consider this 

problem in a future study. 

CONCLUSIONS 

We have compared the methods for estimating the spectral 

reflectance of art paintings for the development of spectral 

color imaging systems. Three traditional methods and three 

methods based on machine learning for spectral reflectance 

estimation of paint were utilized. The traditional methods 

include two linear methods—the method based on PCA and 

the method based on Wiener estimation—and one method 

using multiple regression analysis. We introduced two novel 

machine learning methods utilizing regressive PCA and 

piecewise linear Wiener estimation. Thus, the machine 

learning methods include two methods working with a global 

nonlinear data structure—the method based on regressive 

PCA and the method based on regression analysis—and 

the method using piecewise linear Wiener estimation. Similarly 

to the submanifold method, 20 the learning methods 

used fall between the linear and Bayesian approaches, and 

the method for working with nonlinear data have a limitation. 

They work only with a data structure with a one-toone 

mapping. Finally, we synthesized a spectral color imaging 

system implementing the different estimation methods 

and demonstrated the possibility for accurately estimating 

the reflectance spectra using the presented techniques. 

ACKNOWLEDGMENT 

The authors thank the Academy of Finland for the funding 

granted to this study. 

REFERENCES 

1 Y. Miyake, “Evaluation of image quality based on human visual 

characteristics”, in Proc. of the First International Workshop on Image 

Media Quality and its Applications (IEICE, Nagoya, Japan, 2005) pp. 

10–14. 

2 Y. Miyake, Y. Yokoyama, N. Tsumura, H. Haneishi, K. Miyata, and J. 

Hayashi, “Development of multiband color imaging systems for 

recording of art paintings”, Proc. SPIE 3648, 218–225 (1999). 

3 H. Haneishi, T. Hasegawa, A. Hosoi, Y. Yokoyama, N. Tsumura, and Y. 

Miyake, “System design for accurately estimating the spectral reflectance 

of art paintings”, Appl. Opt. 39, 6621–6632 (2000). 

4 N. Tsumura, H. Haneishi, and Y. Miyake, “Estimation of spectral 

reflectances from multi-band images by multiple regression analysis”, 

Jpn. J. Opt. 27, 384–391 (1998). 

5 M. J. Vhrel and H. J. Trussel, “Color correction using principal 

components”, Color Res. Appl. 17, 328–338 (1992). 

6 M. J. Vrhel and H. J. Trussel, “Filter considerations in color correction”, 

IEEE Trans. Image Process. 3, 147–161 (1994). 

7 S. Goodall, P. H. Lewis, K. Martinez, P. A. S. Sinclair, F. Giorgini, M. J. 

Addis, M. J. Boniface, C. Lahanier, and J. Stevenson, “SCULPTEUR: 

Multimedia Retrieval for Museums”, Proc. of the International 

Conference Image and Video Retrieval (CIVR, Dublin, Ireland, 2004) pp. 

638–646. 

8 K. Martinez, J. Cupitt, and D. Saunders, “High resolution colorimetric 

imaging of paintings”, Proc. SPIE 1901, 25–36 (1993). 

9 J. E. Farrell, J. Cuppitt, D. Saunders, and B. A. Wandel, “Estimating 

spectral reflectances of digital images of art”, Proc. of the International 

Symposium of Multispectral Imaging and Color Reproduction for Digital 

Archives (Chiba University, Chiba, Japan, 1999) pp. 58–64. 

10 J. Y. Hardeberg, H. Brettel, and F. Schmitt, “Spectral characterization of 

electronic cameras”, Proc. SPIE 3409, 100–109 (1998). 

11 H. Maître, F. Schmitt, J.-P. Crettez, Y. Wu, and J. Y. Hardeberg, 

“Spectrophotometric image analysis of fine art paintings”, Proc. IS&T/ 

SID 4th Color Imaging Conference (IS&T, Sprigfield, VA, 1996) pp. 

50–53. 

12 M. Hauta-Kasari, K. Miyazava, S. Toyooka, J. Parkkinen, and T. 

Jaaskelainen, “Spectral vision system based on rewritable broad band 

color filters”, Proc. of the International Symposium of Multispectral 

Imaging and Color Reproduction for Digital Archives (Chiba University, 

Chiba, Japan, 1999) pp. 155–158. 

13 P. D. Burns and R. S. Berns, “Analysis of multispectral image capture”, 

Proc. IS&T/SID 4th Color Imaging Conference (IS&T, Springfield, VA, 

1996) pp. 19–22. 

14 F. H. Imai and R. S. Berns, “High-resolution multi-spectral image 

archives: A hybrid approach”, Proc. IS&T/SID 6th Color Imaging 


15 L. T. Maloney, “Evaluation of linear models of surface spectral 

reflectance with small number of parameters”, J. Opt. Soc. Am. A 10, 

1673–1683 (1986). 

16 J. Parkkinen, J. Hallikainen, and T. Jaaskelainen, “Characteristic spectra 

of Munsell color”, J. Opt. Soc. Am. A 6, 318–322 (1989). 

17 M. J. Vrhel, R. Gershon, and L. S. Iwan, “Measurement and analysis of 

object reflectance spectra”, Color Res. Appl. 19, 4–9 (1994). 

18 A. Ribes, F. Schmitt, and H. Brettel, “Reconstructing spectral reflectances 

of oil pigments with neural networks”, in Proc. of the Third International 

Conference on Multispectral Color Science (University of Joensuu, 

Joensuu, Finland, 2001) pp. 9–12. 

19 D. H. Brainard and W. T. Freeman, “Bayesian color constancy”, J. Opt. 

Soc. Am. A 14, 1393–1411 (1997). 

20 J. M. DiCarlo and B. A. Wandell, “Spectral estimation theory: Beyond 

linear but before Bayesian”, J. Opt. Soc. Am. A 20, 1261–1270 (2003). 

21 I. T. Nabney, Netlab Algorithms for Pattern Recognition (Springer, Berlin, 

2002). 

22 Netlab Toolbox, http://www.ncrg.aston.ac.uk/netlab. 

23 V. Bochko and J. Parkkinen, “Principal component analysis using 

approximated principal components”, Research Report 90, Department 

of Information Technology, Lappeenranta University of Technology 

(2004) pp. 1–7. 

24 A. Corduneanu and C. M. Bishop, “Variational Bayesian model selection 

for mixture distributions”, Proc. of the Eighth International Conference 

on Artificial Intelligence and Statistics, editedbyT.RichardsonandT. 



Jaakkola (Morgan Kaufmann, 2001) pp. 27–34. 

25 J. Tajima, M. Tsukada, Y. Miyake, H. Haneishi, N. Tsumura, M. 

Nakajima, Y. Azuma, T. Iga, M. Inui, N. Ohta, N. Ojima, and S. Sanada, 

“Development and standardization of a spectral characteristics data base 

for evaluating color reproduction in image input devices”, Proc. SPIE 

3409, 42–50 (1998). 

26 B. Funt, D. Kulpinski, and V. Cardei, “Non-linear embeddings and the 

underlying dimensionality of reflectance spectra and chromaticity 

histograms”, Proc. IS&T/SID Ninth Color Imaging Conference (IS&T, 





Digital Watermarking of Spectral Images Using PCA-SVD 

Long Ma 

School of Information Science & Engineering, Northeastern University, Shenyang 110004, China 

Changjun Li 

Department of Color Chemistry, University of Leeds, Leeds LS2 9JT, United Kingdom 

E-mail: C.Li@leeds.ac.uk 

Shuni Song 

School of Science, Northeastern University, Shenyang 110004, China 

Abstract. Kaarna et al. [J. Imaging Sci. Technol. 48, 183–193 

(2004)] presented a technique based on principal component analysis 

(PCA) to embed a digital watermark containing copyright information 

into a spectral image. In this paper, a hybrid watermarking 

method based on the pure PCA approach of Kaarna et al. and singular 

value decomposition (SVD) is proposed. The performance of 

the proposed technique is compared with a pure PCA based technique 

against attacks including lossy image compression, median, 

and mean filtering. The experiments show that the proposed method 

outperforms a pure PCA based technique. 



Received Apr. 21, 2006; accepted for publication Oct. 23, 2006. 

1062-3701/2007/511/79/7/$20.00. 

INTRODUCTION 

Copyright protection is becoming more important in open 

networks since digital copying maintains the original data 

and copying can be easily made at high speed. Digital watermarks 

offer a possibility to protect copyrighted data in the 

information society. A watermarking procedure consists of 

two parts: watermark embedding and extraction algorithms. 

The watermarking procedure should maintain the following 

properties: the watermark should be undetectable and hidden 

from an unauthorized user, the watermark should be 

invisible or inaudible in the information carrying signal, 

and, finally, the watermark should be robust towards possible 

attacks. 1,2 

Normal RGB images have three color bands and the 

information for those bands is integrated from the wavelengths 

of visible light. The spectral images have a large 

number of bands and they may contain information from a 

wider spectrum, also outside the visible range. Spectral imaging 

has various applications in remote sensing and can 

now be used in industrial applications including quality control, 

exact color measurement, and color reproduction. This 

evolution has been possible due to the development in the 

spectral imaging systems. 3–5 

Several watermarking techniques have been developed 

for spectral images. The grayscale watermark can be embedded 

in the transform domains such as by discrete wavelet 

transform (DWT) of the spectral image. 6,7 Furthermore, the 

grayscale watermark can also be embedded in the principal 

component analysis (PCA) transform domain of the spectral 

image. 5,8 

In this paper, a new watermarking method is proposed 

for spectral imaging. The paper is organized as follows: principal 

component analysis (PCA) and singular value decomposition 

(SVD) are briefly discussed in the next section. The 

proposed watermarking procedure is described in the third 

section. The performance measures for the watermarking 

techniques are introduced in the fourth section. Comparisons 

with the pure PCA approach of Kaarna et al. 5,8 are 

given in the fifth section and the conclusions are drawn in 

the final section. 

PCA AND SVD 

In order to describe the new proposed method, the principal 

component analysis (PCA) and singular value decomposition 

(SVD) are briefly discussed here. 

Principal Component Analysis 

PCA has been widely applied to spectral image analysis and 

spectral image coding. 5,9,10 Let =x be a given data set 

containing N column vectors. The main features of the data 

set can be extracted using the PCA algorithm, which has 

the following three steps: 

(1) Compute the mean of the data set: 

= 1 x. 1 

N x 

(2) Compute the covariance matrix C defined by 

C = 1 x − x − T . 

N x 

(3) Compute eigenvalues i and eigenvectors s i of the 

symmetric and semi-positive definite matrix C with 

1 2 ¯ n 0. Heren is the number of components 

2 

79

Ma, Li, and Song: Digital watermarking of spectral images using PCA-SVD 

of the vector x x, and normally nN. Superscript T is 

the transpose of a vector or matrix. 

The vectors s i i=1,2,...,n form a basis for space R n 

(the set of column vectors with n components). Thus, for 

any x, 

n 

x = x T s i s i . 

i=1 

In general, a smaller integer pn can be chosen so 

that the first p eigenvectors’ combination is a good approximation, 

i.e., 

p 

x x T s i s i . 

i=1 

Singular Value Decomposition 

Every mm real matrix A has the following decomposition: 

A = UV T , 

where U and V are mm orthogonal matrices, respectively, 

and is a diagonal matrix having the following form: 

3 

4 

5 

1 0 

= diag 1 , ..., m . 6 

0 m= 

Here i are the singular values, and they satisfy 

1 2 ¯ m 0. 

The above decomposition is called the singular value 

decomposition (SVD) of the matrix A. Ifweletu i and v i be 

the column vectors of the orthogonal matrices U and V, 

then 

m 

A = i u i v T i . 

i=1 

Note that the SVD separates A into two parts: U, V similar 

to the eigenvectors or components in the PCA, and similar 

to the eigenvalues in PCA. Hence, the main components of 

A are u i v i T , and i decides the proportion of those main 

components u i v i T . 

Note also that in this paper it is assumed that when 

PCA or SVD is applied, the resulting eigenvalues i or singular 

values i are arranged in descending order of inequality 

(7). In addition, symbol A ij denotes the i,j position 

element or pixel of the matrix or image A. 

THE PROPOSED WATERMARKING PROCEDURE 

Consider an mm spectral image having n bands. Thus 

each band image S k can be represented as a matrix of the 

following form: 

7 

8 

k 

11 

S =s k 

¯ 

k 

s 1m 

] ] ] k =1,2, ...,n. 9 

¯ 

k, 

k 

s m1 

s mm 

Let be the set of the spectral vectors r ij defined by 

r T ij = s 1 ij ,s 2 ij , ...,s n ij , i,j =1,2, ...,m. 10 

Thus the set has N=m 2 spectral vectors. In addition, we 

assume the watermark image W is a mm gray scale image. 

Now we are ready to describe the proposed watermarking 

procedure: 

Watermark Embedding Algorithm 

(1) Apply SVD to the visual watermark image W, resulting 

in 

with 

W = U w w V w T , 

w = diag w,1 , ..., w,m . 

11 

12 

(2) Apply PCA to spectral domain of the spectral image, 

i.e., apply PCA to the set , resulting in eigenvalues i and 

eigenvectors s i for i=1,2,...,n. Thus, for each r ij in we 

have 

n 

r ij = r T ij s k s k . 

k=1 

13 

Hence, for each k, k=1,2,...,n the kth eigenimage E k 

can be defined by letting E k ij be r ij T s k , i.e., E k ij =r ij T s k 

with i,j=1,2,...,m. Thus, the spatial size of each eigenimage 

E k is the same as that of the original spectral image. 

(3) Choose an eigenimage E k with k satisfying 

1kn and apply SVD to it, resulting in 

with 

E k = U e e V e 

T 

e = diag e,1 , ..., e,m . 

14 

15 

(4) Modify the singular values of the eigenvalue image 

E k by mixing the singular values of the watermark image 

W with those of E k : 

e,1 

¯ e,i = e e,i + w w,i , i =1,2, ...,m, 16 

w,1 

where the coefficients e and w control the strength of the 

embedding. 

(5) Obtain the modified eigenimage: 

Ē k = U e ¯ eV e T , with ¯ e = diag¯ e,1 , ...,¯ e,m . 17 

(6) The watermarked spectral image is constructed by 

computing the modified spectral vectors r¯ij 



i,j=1,2,...,m using the inverse PCA transform, the 

modified eigenimages, and the original spectral eigenvectors. 

Let 

F =Ek , if k k 

k 18 

Then 

n 

Ē k , if k = k. 

r¯ij = F k ij s k , i,j =1,2, ...,m, 19 

k=1 

where s k ’s are the eigenvectors obtained in step (2). 

Note that the watermark embedding algorithm given by 

Kaarna et al. 5 is simpler than the one given above. Their 

algorithm does not need the singular value decompositions 

of the watermark image W [Eq. (11)] and kth eigenimage 

(or multiplier image) E k [Eq. (14)]. Their watermark embedding 

can be simply expressed by 

Ē k = E k + W. 

20 

Thus, their method hides the whole watermark image into 

the original image. However, the watermark image W may 

consist of fine detail over a significant portion of a slowly 

varying background level. Hence the gray levels of W often 

change rapidly around the edges. Thus, the change from 

E k to Ē k [Eq. (20)] is inevitably not “smooth,” which 

will result in greater visual difference between the original 

and the watermarked images, i.e., the watermark will be 

more visible. On the other hand, by using the singular values 

decomposition W=U w w V T w , the watermark image is thus 

separated into two parts: main components U w , and V w or 

v w,i v T w,i ; and w or w,i . Our proposed method only hides w 

into the original image. Since W has m 2 values (pixels), 

while w has only m values, it is expected that our proposed 

method has a better embedding quality. In fact, from Eqs. 

(11) and (14) the modified eigenimage Ē k defined by Eq. 

(17) can be expressed by 

Ē k = U e ¯ eV e T = e U e e V e T + U e w V e T . 

21 

Thus, the proposed change from E k to Ē k is expected to 

be “smooth” (avoiding sudden changes between pixels), 

which will result in less visual difference between the original 

and watermarked images. 

Note also that naturally the watermarked spectral image 

differs from the original image. The difference between these 

images depends on the strength e and w in the watermark 

embedding according to Eq. (16). The strength balances between 

properties like robustness, invisibility, security, and 

false-positive detection of the watermark. The selection of a 

band or the integer k in step (3) affects the visibility of the 

watermark in the watermarked image and the quality of the 

reconstruction of the watermark image against possible attacks. 

It is clear that the smaller the value of k, the more 

visible the watermark will be. On the other hand, the larger 

the value of k, the lower resistance against attacks. All these 

effects will be investigated below. 

Watermark Image Extraction Algorithm 

(1) Compute e ij =r¯ij T s k , with k being defined in step 

(3) of the watermark embedding algorithm, and form matrix 

or image E by setting E ij =e ij for 

i,j=1,2,...,m. 

(2) Apply SVD to E, resulting in 

E = UV T with = diag 1,..., m . 

Note that the singular value i is or is approximately equal 

to ¯ e,i in Eq. (16) 

(3) Reconstruct or estimate the singular values of the 

watermark image by inversing Eq. (16): 

(4) Reconstruct the watermark image W r by computing 

¯ w,i = i − e e,i 

w 

e,1 

w,1 

, i =1,2, ...,m. 

r 

T 

W PCA+SVD = U w ¯ wV w 

with ¯ w = diag¯ w,1 , ...,¯ w,m . 

Note that the above extraction algorithm needs some 

information from the embedding algorithm. They are the 

kth eigenvector s k generated from step (2), the watermark 

image’s decomposition matrices U w , w , V w in Eq. (11), the 

singular values e of eigenimage E k , obtained in step (3), 

and the strength e and w . Hence they should be kept by 

the owner. However, if storage space is critical, U w , w , V w 

need not be kept since the watermark image should be available 

to the owner, hence the values can be obtained by singular 

decomposition when needed. 

Note also that any attack (filtering or compression) on 

the watermarked image will degrade the reconstructed watermark 

image. If the pure PCA approach is used for the 

watermark embedding, any attack will somehow directly affect 

the whole watermark image or U w , w , V w . Thus, the 

reconstructed watermark image is given by 

r 

W PCA = Ū w ˆ wV¯ T w. 

While if our embedding method is used, an attack will only 

affect w ,or w,i since the main components U w , V w of the 

watermark image are not hidden in the watermarked image. 

Hence it is expected that the reconstructed watermark image 

is given by 

r 

W PCA+SVD = U w ¯ wV T w , 

which is closer to W=U w w V T r 

w than W PCA 

=Ū w ˆ wV¯ w T ,ifˆ w 

and ¯ w are not too different, which is confirmed by numerical 

simulations given in the fifth section. 

QUALITY MEASUREMENTS OF THE EMBEDDING 

AND EXTRACTION 

The watermark must be not only imperceptible but also robust 

so that it can survive some basic attacks and image 

distortions. Since spectral images are often very large in both 



spectral and spatial dimensions, lossy image compression is 

usually applied to them. In general, however, lossy compression 

will lower the quality of the image and of the extracted 

watermark. 3D-SPIHT is the modern-day benchmark for 

three-dimensional image compression. 11 

Other possible attacks are different kinds of filtering operations, 

such as median filtering and mean filtering. 

The quality in embedding was measured by peak signalto-noise 

ratio (PSNR), which is defined as 

PSNR = 10 log 10 

nm 2 s 2 

E d , 22 

where E d is the energy of the difference between the original 

and watermarked images, n is the number of bands in the 

spectral image, m 2 is the number of pixels in the image, and 

s is the peak value of the original spectral image. 

Note that the closer the original and watermarked spectral 

images, the smaller the value of E d . Hence, the larger the 

value of PSNR, the closer will be the original and watermarked 

spectral images. 

Another performance measure 5,8 used is to test the average 

spectra for each band of the watermarked spectral image 

and to compare with the average spectra for the corresponding 

band of the original spectral image. The smaller 

difference indicates the better performance of the watermark 

embedding technique. The large changes may induce incorrect 

results as, for example, in classification applications. 

Kaarna et al. 5 used the correlation coefficient for measuring 

the similarity between the original and reconstructed 

watermark images. In this paper, the correlation coefficient 

(cc) is also used as a measure of the quality of the extracted 

watermark image and is defined as 

cc = 

 

m 

 

i=1 

m 

m 

 

i=1 j=1 

m 

 

j=1 

W ij 

− ¯ W r ij 

− ¯ r 

m 

m 

 

2 W ij 

− ¯ i=1 j=1 

W r ij 

− ¯ r 2 

, 

23 

where W and W r are the original and reconstructed watermark 

images, respectively, and where ¯ and ¯ r are the 

mean values of the gray level values of the original and reconstructed 

watermark images, respectively. Note that the 

measure cc is equal to unity if the two images W and W r 

are the same. Hence, the closer to 1 the measure cc is, the 

closer to the original watermark image W the extracted watermark 

image W r will be. 

Besides, the root-mean-square (RMS) error defined by 

Eq. (24) below is also used as a similarity measure between 

the original and reconstructed watermark images: 

Figure 1. a The watermark image: b The band 30 image from the 

BRISTOL image. 

RMS = 1 m m 

 

m 2 i=1 

W ij − W r ij 2 . 

j=1 

24 

Thus, if RMS=0, W and W r have no difference. Hence, the 

smaller the value of RMS, the better the embedding method. 

SIMULATIONS AND COMPARISONS 

In this section, the proposed method is compared with the 

pure PCA technique. 5,8 The spectral image used in this comparison 

is the BRISTOL 12 image. The spectral range of the 

original BRISTOL image was in the human visual region. 

The image had 128 rows and 128 columns m=128 with 

8 bit resolution and had 32 n=32 bands. The watermark 

was an 8 bit gray scale image having 128128 in spatial 

dimension. The watermark used in this experiment is shown 

in Fig. 1(a) and the band 30 image of the BRISTOL spectral 

image is shown in Fig. 1(b). 

The parameters e and w in Eq. (16) control the 

strength of the watermark embedding. e =1 and w =0 

mean that there was no watermark information embedded. 

When e =0 and w = w,1 / e,1 , purely watermark information 

is embedded. In this study, we set e =0,hence,the 

parameter w controls the strength of the embedding. Figure 

2 shows the difference measured in terms of PSNR between 

the original and watermarked spectral images versus the 

band k in which the watermark information was embedded. 

The value in the horizontal axis is the band k, which 

varies from 1 to 32. The values in the vertical axis are the 

corresponding differences in terms of PSNR computed using 

Eq. (22). The value w =7 was used. From this diagram it 

can be seen that PSNR value increases with the increase of 

the value k, which means that when the watermark information 

was embedded in the smaller band k, the energy E d 

of the difference between the original and watermarked 

spectral images is larger, hence the watermark is more perceptible. 

While when the watermark information was embedded 

in the larger band k, the energy E d of the difference 

between the original and watermarked spectral images is 

smaller, hence the watermark is less perceptible. This reflects 

the main characteristics of the PCA transform. The first few 



Table I. Correlation coefficients between the original and extracted watermark images 

using median and trimmed mean filters for the proposed and pure PCA methods. 

PCA+SVD 

Pure PCA 

Median 33 0.999 0.946 

Median 55 0.998 0.916 

Trimmed mean 33 0.995 0.948 

Trimmed mean 55 0.984 0.923 

Table II. RMS error between the original and extracted watermark images using 

median and trimmed mean filters for the proposed and pure PCA methods. 

Figure 2. The difference in terms of PSNR vertical axis between the 

original and watermarked spectral images versus the band k horizontal 

axis in which the watermark information was embedded. 

PCA+SVD 

Pure PCA 

Median 33 10.730 010 1 20.839 572 5 

Median 55 15.438 597 1 25.182 329 5 

Trimmed mean 33 13.611 175 5 20.057 059 7 

Trimmed mean 55 20.874 862 6 24.133 671 1 

Figure 3. Average spectra vertical axis of the original marked * , watermarked 

images using the pure PCA technique marked with and 

using the proposed method marked with + versus band k horizontal 

axis. 

eigenvectors carry the main feature of the spectral set , 

while the additional eigenvectors share little main features of 

the spectral set. 

Embedding Quality Comparison 

We now compare the embedding quality using our proposed 

technique and a pure PCA-based watermarking technique. 

Note that the performances of both techniques depend on 

the choice of the band k and embedding strength parameters. 

In this experiment k was fixed to a value of 3 for both 

methods. For a fair comparison, the strength parameter for 

each method was adjusted so that the difference between the 

original and watermarked spectral images measured in terms 

of PSNR computed using Eq. (22) is approximately equal to 

a given value (PSNR=34.50 dB in this experiment). 

Since each of the embedding methods has the same 

value of PSNR, the comparison of the embedding quality is 

made based on the measure of average spectra. 5,8 Figure 3 

shows the average spectra from the original image (curve 

with “”), the watermarked images using the PCA (curve 

with “”) and the proposed methods (curve with “”) versus 

band k. The vertical values are the average spectral values, 

and the value in the horizontal axis is the band k varying 

from 1 to 32. From this diagram, it can be seen that the 

average spectra curve for the watermarked image embedded 

using the proposed method is nearly overlapping with that 

of the original image, while the average spectra curve of the 

watermarked image using the purely PCA method is markedly 

different from that of the original image at two end 

bands (bands 1–15, and bands 25–32). This diagram clearly 

shows the proposed method is better than the pure PCA 

approach. 

Reconstruction Quality Against Attacks 

It is often that image processing techniques can be applied to 

the watermarked spectral image. The processed or attacked 

image will affect the quality of the reconstructed watermark 

image. The difference between the original watermark image 

W and the extracted watermark image W r measures the 

performance of the watermark embedding method. The 

smaller the difference, the better the method performs, and 

the method is more robust against attack. Also, the small 

difference indicates the closeness between the two images. 

Here the correlation coefficient cc defined by Eq. (23) and 

RMS defined by Eq. (24) are used as measures of the closeness 

between W and W r , or measures as the robustness of 

the watermark embedding method. Median filtering and 



Figure 4. Correlation coefficients between the original and extracted watermark 

images for the proposed and pure PCA methods versus the compression 

ratio using the 3-D SPIHT lossy compression method. 

Figure 5. RMS errors between the original and extracted watermark images 

for the proposed and pure PCA methods versus the compression 

ratio using the 3-D SPIHT lossy compression method. 

trimmed mean filtering were first applied to the watermarked 

image with the filter size being 33 and 55. The 

correlation coefficient (cc) and root-mean-square (RMS) results 

are listed in Tables I and II respectively. In each table, 

the values in the second column indicate the performance or 

robustness of the proposed method PCA+SVD, and the 

values in the last column reflect the performance of the pure 

PCA approach. For example, from the second row of Table I, 

the correlation coefficients were computed after the watermarked 

image was processed using a 33 median filter. The 

correlation coefficient (second row and column) for the proposed 

method is 0.999, while the correlation coefficient (second 

row and third column) for the pure PCA approach is 

0.946. In fact, when comparing the results in each row of 

Table I, it can be seen that values in the second column are 

always closer to a value of unity than the corresponding 

values in the last column, indicating the proposed method 

outperforms the pure PCA approach. Similarly, from Table 

II, the values in the second column are always smaller than 

the corresponding values in the last column, indicating once 

again the proposed method outperforms the pure PCA 

approach. 

Lossy compression was the second attack that corrupted 

the watermarked image. A 3D-SPIHT 11 compression 

method was used. The correlation coefficient and RMS error 

(vertical axis) between W and W r versus the compression 

ratio (horizontal axis) are shown in Figs. 4 and 5, respectively. 

The curve marked with “ * ” corresponds to the proposed 

method PCA+SVD and the curve marked with “” 

corresponds that to the pure PCA approach. It can be seen 

from Fig. 4, that the curve marked with “ * ” is much higher 

than the curve marked with “,” telling us that for each of 

the chosen compression ratios, the correlation coefficient for 

the proposed method is closer to unity than that for the pure 

PCA approach. Figure 5 shows that the RMS error curve of 

the proposed method is always located below that of the 

pure PCA approach. Both measures show the proposed 

method is more robust than the pure PCA method. 

CONCLUSIONS 

In this paper, a digital watermarking technique has been 

proposed based on principal component analysis (PCA) and 

singular value decomposition. This work was motivated by 

the pure PCA approach of Kaarna et al. 5,8 The robustness of 

the proposed method was tested using attacks such as lossy 

compression, median, and trimmed mean filtering. Simulation 

results have shown the proposed method is more robust 

than the pure PCA approach. However, comparing with the 

PCA approach our method involves further singular value 

decompositions of the watermark image and selected band 

of eigenimage E k , which will give no problems with modern 

computing power considering the nature of the 

applications. 


The authors are indebted to Mike Pointer for his valuable 

suggestions, which improved the quality of the paper. 

REFERENCES 

1 D. Arta, “Digital steganography: Hiding data within data”, IEEE Internet 

Comput. 5, 75–80 (2001). 

2 C. I. Podilchuk and E. J. Delp, “Digital watermarking: Algorithms and 

application”, IEEE Signal Process. Mag. 18, 33–46 (July 2001). 

3 M. Hauta-Kasari, K. Miyazawa, S. Toyooka, and J. Parkkinen, “Spectral 

vision system for measuring color images”, J. Opt. Soc. Am. A 16(10), 

2352–2362 (1999). 

4 T. Hyvärinen, E. Herrala, and A. Dall’Ava, “Direct sight imaging 

spectrograph: A unique add-on component brings spectral imaging to 

industrial applications”, Proc. SPIE 3302, 165–175 (1998). 

5 A. Kaarna, P. Toivanen, and K. Mikkonen, “PCA transform in 

watermarking spectral images”, J. Imaging Sci. Technol. 48, 183–193 

(2004). 

6 A. Kaarna and J. Parkkinen, “Digital watermarking of spectral images 

with three-dimensinal wavelet transform”, Proceedings of the 

Scandinavian Conference on Image Analysis, SCIA 2003 (Springer, 

Halmstad, Sweden, 2003) pp. 320–327. 

7 A. Kaarna and J. Parkkinen, “Multiwavelets in watermarking spectral 



images”, Proceedings of the International Geoscience and Remote Sensing 

Symposium, IGARSS’04 (IEEE, Piscataway, NJ, 2004) pp. 3225–3228. 

8 A. Kaarna, V. Botchko, and P. Galibarov, “PCA component mixing for 

watermarking embedding in spectral images”, Proc. IS&T’s 2nd 

European Conference on Color in Graphics, Imaging, and Vision 

(CGIV’2004) (IS&T, Springfield, VA, 2004) pp. 494–498. 

9 A. Kaarna and J. Parkkinen, “Transform based lossy compression of 

multispectral images”, Pattern Anal. Appl. 4, 39–50 (2001). 

10 A. Kaarna, P. Zemcik, H. Kälviäinen, and J. Parkkinen, “Compression of 

multispectral remote sensing images using clustering and spectral 

reduction”, IEEE Trans. Geosci. Remote Sens. 38, 1073–1082 (2000). 

11 P. L. Dragotti, G. Poggi, and A. R. P. Ragozini, “Compression of 

multispectral images by three-dimensional SPIHT algorithm”, IEEE 

Trans. Geosci. Remote Sens. 38, 416–428, 2000. 

12 BRISTOL, http://www.crs4.it/~gjb/ftpJOSA.html, accessed 10 October 

1998. 




Qualification of a Layered Security Print Deterrent 1 

Steven J. Simske and Jason S. Aronoff 

Hewlett-Packard Laboratories, 3404 E. Harmony Rd., Mailstop 85, Fort Collins, CO 80528 

E-mail: Steven.Simske@hp.com 

Abstract. Variable data printing (VDP), combined with precision 

registration of multiple ink layers, empowers a layered deterrent using 

variable print strategies on each of the multiple layers. This shifts 

the need for specialized printing techniques to the need to accommodate 

variable ink approaches. Such layered deterrents can incorporate 

infrared/ultraviolet fluorescent inks, infrared opaque and 

transparent black inks, inks containing taggants, magnetic ink, and 

inks with differential adhesive properties to enable sandwich printing. 

Overt features printed as part of the same layered deterrent 

provide excellent payload density in a small printed area. In this 

paper, the statistical and hardware processes involved in qualifying 

two layers of such a deterrent for their deployment in product (e.g., 

document and package) security are presented. The first is a multicolored 

tiling feature that provides overt security protection. Its color 

payload is authenticated automatically with a variety of handheld, 

desktop, and production scanners. The second security feature is 

covert and involves the underprinting or overprinting of infrared information 

with the covert tiles. Additional layers using existing security 

deterrents are also described, affording the user information 

densities as high as 560 bits/cm 2 70 bytes/cm 2 . 



INTRODUCTION 

Counterfeiting, smuggling, warranty fraud, production overruns, 

product diversion, and related problems are a huge 

concern for brand owners. Conservative estimates place 

counterfeiting alone at 5–7% of world trade, or more than 

$300 billion/annum. 1 Because the harmful effects of counterfeiting 

extend to entire economies and societies, 2 fighting 

counterfeiting not only protects a brand name but also can 

add to brand value if the company is perceived as an agent in 

product security. Counterfeiting in the pharmaceutical industry 

is enabled by the practice of relabeling and 

repackaging, 3 increasing the need for item-level authentication. 

The US Food and Drug Administration (FDA) has 

created a Medwatch 4 program to provide up-to-the-minute 

reporting of adverse events in the pharmaceutical distribution 

chain, emphasizing the ubiquity and severity of the 

counterfeiting. 

To deter counterfeiters, a layered deterrent is recommended. 

This is a printed deterrent that contains two or 

more layers of information in a single region. Higher density 

of layered deterrents is provided when multiple layers of ink 

1 Presented in part at IS&T’s Digital Fabrication Conference, Baltimore, 

MD, September, 2005. 

Received Jan. 25, 2006; accepted for publication Aug. 15, 2006. 

1062-3701/2007/511/86/10/$20.00. 

are precisely registered, such as is possible with liquid electrophotographic 

(LEP) digital press technologies. 

Product security begins with the package. If each package 

provides a unique identifier, which can be tracked and 

linked to a provenance record tracing its location throughout 

its distribution path, then even a modest level of 

customer/retailer authentication poses a significant exposure 

risk to a would-be counterfeiter. 5 The incentive for package 

reuse is also removed. Using this approach, the packages 

should provide overt security printing features that can be 

authenticated simply (e.g., with camera phones, digital cameras, 

scanners, and all-in-ones) and reliably. This approach 

will always be complemented by complex deterrents (colorshifting 

inks, layered deterrents, 6 etc.), electronic and active 

deterrents (RFID, etc.), tamper-evident deterrents, and other 

registry-based deterrents. Under some circumstances, a 

unique identifier can provide a level of security dictated by 

its density—the amount of information that can be reliably 

read using the deterrent. For this to happen, it must be 

reliably authenticated. 

In this paper, two deterrents are considered (and salient 

portions of them qualified). The first is a 2D arrangement of 

color tiles, 7,8 which can provide branded colors, productspecial 

colors, and/or be part of an overt deterrent. These 

color tiles can in turn be associated with overprinted microtext. 

Figure 1 demonstrates this feature in its two default 

deployments: without superimposed microtext (upper) and 

with superimposed microtext (lower). The upper color tile 

feature can also accommodate hidden ultraviolet/infrared 

(UV/IR) inks, as described below—or overprinted UV/IR 

inks—for additional, covert security. In Fig. 1, the default 

deployment of the upper feature is expanded to twice its size 

relative to the lower feature (the addition of microtext requires 

roughly a 2 increase in tile width and height to 

authenticate accurately). 

Thirty-six characters (the 26 English letters A–Z and the 

10 numerals 0–9) are associated with two consecutive color 

tiles (each taking on one of six possible colors—thus, the 36 

characters are encoded exactly by 66 color combinations) 

in English reading order (left to right by row, top to bottom 

by consecutive rows). The color pairs mapped to these are 

A=R,R, B=R,G, C=R,B, D=R,C, E=R,M, 

F=R,Y, G=G,G..., 9=Y,Y, whereRGBCMY are the 

colors red, green, blue, cyan, magenta, and yellow, respectively. 

Note that, for example, the letter “N” is always encoded 

as a blue followed by a green tile in the feature on the 

86

Simske and Aronoff: Qualification of a layered security print deterrent 

Figure 1. Color tile security printing feature in default deployment, without 

microtext upper and with microtext lower. The upper feature is 

expanded to twice its size relative to the lower feature, as necessary for 

accurate authentication the addition of microtext requires approximately 

a twofold increase in tile width and height to authenticate accurately. The 

letters A–Z and numerals 0–9 are associated with two consecutive color 

tiles in European reading order. The color pairs mapped to these are A 

=R,R, B=R,G, C=R,B, D=R,C, E=R,M,...,9=Y,Y, 

where RGBCMY are the colors red, green, blue, cyan, magenta and 

yellow, respectively. Note that, for example, the letter “P” is always encoded 

as a blue followed by a cyan tile in the feature on the right above. 

right in Fig. 1. Both features encode the string “THISWAS- 

PRINTEDFORJOURNALOFIMAGINGSCIENCEAND- 

TECHNOLOGY15JAN2006.” 

The second layer (Fig. 2) is a binary covert tile produced 

by one of two approaches. The first approach is the combination 

of an infrared (IR) reflective ink layer overprinted by 

two types of black (or other spot color) ink, 9,10 making it 

Figure 2. Security printing features: color tile upper and binary tile 

lower for testing differential IR-opaque inks. For the qualification described 

herein, the color tile feature was printed using CMY cyan, magenta, 

yellow inks, and the binary tile feature with spot color blue 

C6170A ink. 

appear to be a uniform (spot) colored area, but encoding a 

covert tile structure. This feature is produced using inks that 

have differential opacity to visible and infrared light excitation. 

In offset and other “static printing” technologies, process 

black ink can be used as the ink with opaque IR characteristics, 

and Anoto black ink 11 as the ink with transparent 

IR characteristics. Using a variable data printing front end, 

one can simply select between the two spot color inks and 

decide which sections of underprinted infrared ink to reveal. 

The second approach to providing a layered deterrent is to 



simply overprint IR tile patterns on a color tile deterrent 

such as shown in Fig. 1. The second approach was simulated 

here with a blue ink tile. 

SECURITY PRINTING FEATURE QUALIFICATION 

To qualify a feature, the following steps are required 

a. Design the feature. This includes specifying the 

variables in the feature and the ranges over which 

they should be varied. On the low end of the 

range, the feature should essentially never authenticate 

(or authenticate below any acceptable accuracy), 

whereas on the high end, the feature should 

authenticate at an acceptable level. For the color 

tile feature, the variables include (i) the set of colors 

printed, (ii) the width and height of, and thus 

number of bits in, the feature, (iii) the inclusion/ 

exclusion of (visible) microtext, and (iv) the width 

and height of the tiles. 

For the binary tile, the variables include (i) the 

spectral characteristics of the inks used, (ii) the 

width and height of, and thus number of bits in, 

the feature, and (iii) the width and height of the 

tiles. 

b. Determine the set of features to print. Based on 

the above set of variables, for the color tiles (i) the 

set of colors printed was RGBCMY, (ii) an 

88 array of tiles was printed with at least six of 

each color, (iii) microtext was not printed visibly 

over the color tiles, and (iv) the width and height 

are equal and are varied from 0.125 to 1.25 mm 

(in 0.125 mm increments). 

For the binary tiles, (i) a single ink was selected 

to print, HP C6170A spot color blue ink, 

(ii) an 89 array of tiles was printed with 32 

white spaces and 40 black spaces (including 8 

black spaces on the lowest row, as in Fig. 2), and 

(iii) the width and height are equal and are varied 

from 0.125 to 1.25 mm in 0.042 mm increments. 

c. Print the set of features. Thirty-six color tile features 

were printed at each of ten sizes, at 600 ppi. 

For purposes of testing, multiple security printing 

features are written to each letter-sized 

118.5 page, as shown in Fig. 3. A total of 360 

(36 each at 0.125, 0.25, …, 1.25 mm in dimension) 

color tile features, each with 60 colored tiles 

(21 600 total tiles), were printed. The final four 

black tiles on the color tile features are ignored by 

the authentication algorithm. The color tile features 

were printed on a thermal inkjet printer at 

600 dots per inch dpi, or240 dots/cm, resolution 

using default settings except for selecting 

“high quality.” 

A total of 16 binary tile features were printed 

at resolutions of 0.125, 0.167, …, 1.25 mm (28 

different sizes, 16 binary tiles each, 72 tiles each, 

for a total of 32 256 tiles). A sample page for these 

tile features is shown in Fig. 3. The binary tile 

features were printed on a thermal ink jet printer 

at 600 dpi 240 dots/cm using default settings, 

except that the color cartridge was disabled (so 

only the blue ink printed) and “high quality” was 

selected. 

Each color tile sequence used 30 of the 36 

characters in the set, and each character appeared 

in 30 of the 36 samples at each resolution (the 

same set of 36 features was printed at each resolution). 

Each binary tile included the 16 4-bit subsequences 

(0000, 0001, …, 1111), and once more 

the same set of 16 features was printed at each 

resolution. 

d. Scan the pages of the features. The printed pages 

were all scanned using a commercial off-the-shelf 

desktop scanner (the pages were placed manually 

on the scanner, so that the automatic document 

feeder was not used) at 600 pixels per inch ppi, 

or 240 dots/cm, using default settings, and stored 

with lossless compression. To accommodate all the 

features, 29 pages of color tiles and 37 pages of 

binary tiles were scanned. 

e. Extract the features from the printed pages. A segmentation 

algorithm 12,13 was used to extract each 

feature automatically from the scanned page of 

multiple features. Where possible, whitespace was 

included around the feature. After this step, the 

360 color tile features and 448 binary tile features 

are saved as individual image files. 

f. Authenticate the features. The set of extracted features 

is then evaluated using the authentication algorithms 

described below. The output of the authentication 

algorithm is a sequence that can be 

directly compared to the intended sequence. The 

number of loci (single tile reading) errors is calculated 

for each feature. 

g. Determine critical point in the authentication 

curves. Curves are then obtained showing the 

number and percentage of tiles read successfully 

along with the absolute number of tiles correctly 

read. From these data, one can recommend the 

security feature deployment parameters (size, in 

the case of the features tested herein). The error 

rate is used to define how many check bits, redundant 

bits, etc., must be added to prevent read 

errors. 

AUTHENTICATION 

Color tile authentication consists of the following steps, all 

of which are embedded in a single executable that performs 

near-real time analysis of an image: 

a. Thresholding. Thresholding is performed on the 

saturation values of the scanned pixels, since the Y 

tiles have similar intensity values to white, and the 

six colors cover much of the hue gamut. Saturation 

is defined as 



Figure 3. Sample pages printed for qualification: color tiles pair on left side, columns a and b and binary 

tiles pair on right side, columns c and d. For each pair, the raster files to be printed are on the left, and the 

scanned pages are on the right. For the binary tiles, the print raster is binary black and white, and the 

C6170A spot color blue ink was printed and scanned using the “black ink” cartridge in the thermal ink jet 

printer. Both sets represent the largest dimension tested 1.251.25 mm tiles. 



Saturation = 255 1 − minR,G,B 

/sumR,G,B. 

1 

The threshold value is determined from the 

moving average-smoothed saturation histogram 

and is the minimum point of the saturation histogram 

above the peaks for black and white (which 

usually overlap) and the next peak (typically for 

blue). 

b. Segmentation. The resulting thresholded image is 

then prepared for segmentation with a sequence of 

thinning (to eliminate speckle noise), fattening (to 

return nonerased regions to their original size), 

and run-length smearing (to prevent gaps in features). 

These preparatory steps are well known for 

2-D segmentation, extending back 25 years. 14 Because 

we are looking for nontext regions, default 

segmentation preparation as described in Ref. 14 is 

used: we then filter out the regions formed based 

on size and aspect ratio (and later histograms) to 

locate the tile features. Next, regions are formed, 

and the set matching the expected size of the security 

printing features is identified and outlined. 

The processing to this point requires less than 

0.5 s on a mid-range laptop computer for a 

1015 cm 2 image, suitable for authentication of a 

single package or document. For a full page (e.g., 

2025 cm 2 cropped image), the same mid-range 

(2 GHz processor clock, 512 MB RAM) laptop requires 

approximately 3sof processing time. 

c. Subsegmentation. These regions are extracted to 

individual files and corrected for skew, if present. 

The features are then sliced into eight columns and 

eight rows (per the specification of the features as 

88 tiles in size) and these 64 regions assigned in 

reading order. The four black tiles at the end are 

used to make sure they are oriented properly, and 

then discarded to leave a 60 tile sequence. Because 

these images are now the size of the deterrent itself, 

the subsequent steps are performed very rapidly 

(a much smaller image is processed much 

more quickly), generally in less than 10 ms, for 

example, on a mid-range laptop. 

d. Find color peaks. The (CMY) color peaks are 

found first. Separate C, M, and Y maps the same 

size as the feature are created and the values for C, 

M, and Y calculated as 

C = B + G − R, 

M = B + R − G, 

Y = G + R − B. 

Each of these maps is histogrammed, and the 

largest peak above the midpoint (255 for an 8–bit/ 

channel or 24–bit image) of the range (0–511 for 

24-bit image) of the histogram is defined as the C, 

M, orY peak in each of these maps. The median 

value in the peak is taken as the representative 

value for each of these three colors. The pixels assigned 

to any of these three peaks are then ignored 

(not added to the histograms) when the RGB 

color peaks are defined. 

The values for R, G, and B are calculated as 

R = 255 + R − B − G, 

G = 255 + G − B − R, 

B = 255 + B − R − G. 

The pixels not assigned to C, M, orY peaks in 

the previous step are now histogrammed. Here, 

the largest peak above the midpoint of the range of 

the histogram is defined as the R, G, orB peak. 

The median value in the peak is taken as the representative 

value for each of these three colors. 

e. Assign color value to every pixel in the feature. 

Next, the distance from the defined (median) value 

of each peak is computed for every pixel, and each 

pixel is assigned a color value corresponding to the 

minimum distance (Fig. 4, middle image). 

f. Assign color value to every tile in the feature. For 

each tile region, the number of pixels assigned to 

each color is summed, and the color with the 

maximum value is assigned to the tile (Fig. 4, right 

image). Ambiguous tiles (wherein the color with 

the maximum value is assigned less than half the 

pixels) are reported. 

g. Report tile sequence. The 60 tile sequence is organized 

into 30 consecutive pairs. These 30 pairs of 

tiles are decoded into a 30 character string which is 

then compared to the intended sequence. Errors 

are listed as single or dual tile errors (the latter 

counts as two “errors”). 

Figure 4 shows the effects of these steps on a scanned 

color tile feature. The output of the authentication is the 

sequence as follows, which is directly compared to the 

printed sequence (in this case, there is no error): 

“FGHIJKLMNOPQRSTUVWXYZ012345678” 

The steps for authenticating binary tiles are similar to 

that for color tiles, though in general simpler: 

a. Thresholding. Thresholding is again performed on 

the saturation values of the scanned pixels. We 

chose a blue ink to provide the “most challenging” 

thresholding test of the set RGBCMY (blue ink 

had the lowest saturation peak of these six peaks). 

The threshold value is again determined from the 

moving average-smoothed saturation histogram. 

b. Segmentation. Segmentation is performed as for 

color tiles. 

c. Subsegmentation. These regions are extracted to 



Figure 4. a Sample color tile feature after being segmented and extracted from the top of column b of Fig. 

3. b White and black pixels assigned to black and individual pixels assigned to one of the color set 

RGBCMY. c Subsegmentation of the color tile feature and the color assignment of each tile. 

individual files and corrected for skew, if present. 

The features are then sliced into nine columns and 

eight rows (per the specification of the features as 

98 tiles in size) and these 72 regions assigned in 

reading order. The eight consecutive black tiles at 

the end are used to make sure they are oriented 

properly, and then discarded to leave a 64-tile sequence. 

d. Assign foreground/background value to every 

pixel in the feature. For each tile region, the number 

of pixels assigned to foreground (blue) is 

summed, and if this number is greater than the 

number assigned to background (white), then the 

tile is assigned to “foreground” (“1” in the sequence). 

Otherwise the tile is assigned to “background” 

(“0”). 

e. Report tile sequence. The 64 tile sequence is recorded, 

which can then be compared to the intended 

sequence. 



Table I. Results for color tile qualification. Read failures correspond to features with 

insufficient color saturation. The number of correct reads is out of a possible 2160 total 

color tiles read at each tile size. 

Tile 

dimension 

mm 

Read 

failures 

% 

Errorless 

reads 

% 

Tile error 

rate 

% 

Correct 

reads no. 

0.13 100.0 0.0 100.0 0 

0.25 69.4 0.0 93.64 42 

0.38 80.6 0.0 43.33 238 

0.50 52.8 33.3 6.18 957 

0.63 69.4 11.1 7.58 610 

0.75 50.0 2.8 11.02 961 

0.88 8.3 16.7 5.05 1880 

1.00 2.8 30.6 2.48 2048 

1.13 2.8 61.1 1.90 2060 

1.25 0.0 75.0 0.74 2144 

QUALIFICATION 

There were several types of errors in reading the color tiles 

(Table I). The first was due to features having insufficient 

color saturation (Table I, second column from left), in which 

the scanned feature had insufficient consistency of saturation 

of the colors to segment as a single region (due to low saturation 

pixels being assigned to the “black” and “white” pixel 

category). Figure 5 illustrates several examples of these features 

(these are 0.250.25 mm 2 tiles). Halftoning likely 

contributed to this phenomenon, since the “additive” colors 

RGB fared more poorly than the subtractive colors CMY. 

The latter correspond more exactly with the ink pigment 

colors, and so are less affected by halftoning. Features that 

segmented incorrectly were simply registered as “read failures,” 

and these occurred for tile dimensions up to 

1.1251.125 mm 2 . 

The second type of failure was an incorrect color assignment 

for a (properly) segmented tile. This is reported as the 

“tile error rate” (Table I, fourth column from left). This 

value dropped to 6.2% at a tile size of 0.500.50 mm 2 , then 

increased again, dropping to 5.0% at 0.880.88 mm 2 . This 

nonlinear behavior for tiles from 0.50 to 0.88 mm in dimension 

may simply be an artifact of the small number of 

pages scanned. If not, it is likely a consequence of the automatic 

subsegmentation approach of the simple authentication 

algorithm deployed for the qualification work presented 

here. Regardless, by the time the tiles were 1.25 mm on a 

side, read failures had dropped to zero, 75% of the features 

were read without a single tile error, and the overall tile error 

rate was less than 1%. Thus, individual tile reading accuracy 

surpassed 99% at this size (Fig. 6). 

The graph for binary tile errors (Fig. 7) was relatively 

well behaved. The smallest two sizes (3 and 4 pixels, or 0.125 

Figure 5. Color tile patterns 0.250.25 mm in size with low print 

quality. Many of the pixels in the colored RGBCMY areas of these 

features are closer in saturation terms to the black peak than to the color 

peaks. Even when these lower-resolution features are segmented correctly, 

there is a high tile reading error rate TableI,Fig.6. 

and 0.167 mm, on a side for each tile) were essentially unreadable, 

with error rates of 50%. By0.208 mm on a side 

(Fig. 7), however, the tiles were readily readable, with an 

error rate just over 10%. The error rate dropped below 1% 

by the time the binary tiles reached 0.630.63 mm 2 in size. 



Figure 6. Color tile authentication accuracy as a function of tile size. 

99% accuracy is achieved by 30 pixels at 600 ppi, or 240 dots/cm, 

or 1.251.25 mm in widthheight. Tiles are squares ranging from 

0.125 3 pixels to 1.25 30 pixels mm in size. 

Figure 7. Binary tile authentication error rate 100%-accuracy as a function 

of tile size. 99% accuracy is achieved by 15 pixels at 600 ppi, or 

240 dots/cm, or 0.625 mm in width/height. 

DISCUSSION 

Performing the qualification of a security printing feature is 

important to ensure that customers retailers, and/or field 

investigators will willingly and consistently perform authentication. 

Of course, this is not simply a technical issue. An 

important means for encouraging compliance is to put in 

place convenient systems for gracefully handling exceptions 

(read failures, periodic authentication, etc.). Another means 

of improving compliance is to largely eliminate “read failures,” 

which, for example, argues for a 1.251.25 mm 2 

color tile for the hardware used in this qualification study. 

The output of qualification is a recommendation for the 

deployment of the feature: its size and density (e.g., how 

many tiles to use and how large the tiles are), the printing 

and reading/scanner hardware to be used, and the purpose 

of the feature. The latter point was not addressed directly in 

this paper, but is directly related to an accuracy curve such as 

that shown in Fig. 6. If the color tile size is selected to be 

“just beyond” the knee of the curve (e.g., a 1212 pixel, or 

0.50.5 mm 2 , color tile is chosen), then the feature can 

provide an anticopying deterrence in addition to the security 

of the sequence itself. If, on the other hand, the size is made 

as large as possible to prevent any “read failures,” then a 

counterfeiter may be able to more readily copy a batch of 

features. (Since copying degrades the features, it will effectively 

move the feature further toward the “knee” of the 

authentication accuracy curve, but the greater reliability of a 

large tile will prevent a large increase in read failures.) Thus, 

smaller tiles perform a function more like that of a copy 

detection pattern 15 (that is, covert), while larger tiles perform 

a function more like that of a bar code (that is, overt). It is 

important to note that even if a counterfeiter can successfully 

copy an overt feature, the presence of a secure (database) 

registry for polling with the for-authentication sequences 

will always discourage wholesale counterfeiting (so 

long as the codes are actually routinely verified by the end 

user—customer, retailer, and/or field inspector). 

Performing the qualification is also an excellent means 

of evaluating the effectiveness of the authentication system 

one is planning to use with a product. In performing the 

experiments above, for instance, it was observed that for 

tile-based deterrents, there are at least two distinct, broad 

classes of errors made during authentication. The first class 

of errors, which are highly dependent on the size of the tiles, 

and thus follows a classic “S curve” such as that shown in 

Figs. 6 and 7, are broadly termed “printing errors.” These 

errors, which are manifest at sizes larger than the individual 

printing dots, are addressed through improving the printing 

technique (e.g., by changing the hardware, such as using a 

device with more precise ink placement) or approach (e.g., 

by eliminating halftoning through the use of six spot color 

inks for the color tiles), or by changing the ink itself (this is 

not an easy prospect, since ink chemistry is constrained by 

the physics of the printing), with varying improvements. It 

should be noted that these print errors (smearing, blotching, 

etc.), if uncorrected, prevent any increased deterrent density 

through magnification. 

The second type of error is the error associated with the 

authentication algorithm itself. For the experiments described, 

relatively simple authentication approaches were 

adopted. Because of this, we were able to make on-the-fly 

changes to these algorithms to reduce the overall error rate. 

For example, during the performance of the binary tile authentication, 

we noted that occasionally the authentication 

algorithm would crop the 98 feature to effectively an 8 

8 feature. This resulted in infrequent occurrences of a significant 

misread of a feature because the algorithm was attempting 

to impose a 98 structure on an 88 matrix. 

Increasing the size of the gap smeared by the run-length 

smearing eliminated this algorithm error. As a second example, 

during the performance of the color tile feature authentication, 

we noticed that finding the subtractive CMY 

color peaks first reduced the overall error rate considerably 

in comparison to finding the additive RGB color peaks 

first. 

Feature qualification focuses on the different aspects of 



the security printing feature to which the overall authentication 

process is sensitive. The size of the feature, as shown 

here, is clearly an important (perhaps the most important) 

factor. However, many other factors are important to consider, 

including the device independence of the authentication. 

Any off-the-shelf version of the scanning hardware 

used for qualification work should perform as well as the 

one used during qualification. Other factors include control 

over the printing process (for example, being able to reduce 

the effects of halftoning significantly improve color tile authentication 

accuracy), the ability to match the printing and 

scanning resolutions (or at least have them be integral multiples 

of each other), and the processing available for authentication. 

For example, if processing power is unlimited, then 

it is advantageous to put much more intelligence into the 

authentication algorithm, including the ability to respond 

adaptively to ink- and other print-related problems that 

might otherwise contribute to tile read errors. One of the 

principal purposes of qualification is to determine where to 

focus one’s energies—on the printing, the scanning, or the 

authentication. 

Based on the results, the color tile feature can be deployed 

using relatively inexpensive thermal ink jet printers 

and desktop scanners for production and authentication, respectively, 

with a bit density of 160 bits/cm 2 . The binary 

tile feature can be deployed at 250 bits/cm 2 . These densities 

assume that a tile read accuracy of 99% is acceptable. 

More generally, however, these bits will be incorporated into 

an overall deterrent, which includes positioning outline 

(akin to those on a 2-D DataMatrix barcode, for example 16 ) 

and error code checking such as the Reed-Solomon 

algorithm. 17 The final density of these tile-based deterrents, 

then, will be on the order of 100 bits/cm 2 using the authentication 

equipment described herein. 

The qualification work is used to recommend a deployment 

size and parameter definition. It is also used to define 

how many check bits, redundant bits, etc. must be added to 

prevent read errors. For example, at a bit density of 

160 bits/cm 2 , 25% of the color tiles will suffer at least one 

tile classification error. This means that the true “asdeployed” 

density of the tile feature will be reduced to incorporate 

error checking tiles. There is a trade-off between 

reducing the size of the tiles (which increases the tile error 

rate) and needing to incorporate more color tiles to provide 

error checking. An ideal tile-based security printing feature 

reaches a consistent low error rate above a certain size, allowing 

the error-checking approach to be reliably deployed. 

In addition, magnification can be used to increase the density, 

though with exacerbation of any print defects (see Fig. 

5). 

Additionally, the overall “ecosystem” in which the tilebased 

security printing features are to be deployed affects the 

selection of parameters in the features. For example, if the 

raw sequences encoded in the color tiles are stored in a 

(sparse) registry such that the odds of a random sequence 

being in the registry are quite low, 5 then low error rates in 

the color tiles can be overcome by using a pattern matching 

approach (such as that employed in bioinformatics) to find 

the best fit in the registry to the (mis-)reported sequence. 

Sequences with too high a number of errors can either be 

rejected as counterfeit, or trigger an event asking the user to 

rescan the feature. Alternatively, if a large volume of deterrents 

are being scanned simultaneously, the packages with 

“read failures” can be manually authenticated, authenticated 

with a more sophisticated scanner, or simply ignored (due to 

the rest of the deterrents successfully authenticating), depending 

on the needs and governance rules for the product 

and its authentication. 

In-house and externally developed security printing features 

are fully qualified using the processes described herein. 

Printing and scanning are performed with the exact hardware 

to be used by consumers, retailers, and field inspectors. 

In most cases, this will require a plurality of scanning hardware; 

for example, a camera phone or PDA-like device for 

consumers, handheld scanners for retailers, desktop scanners 

for field inspectors, and a vision system for forensic investigators. 

Additional authentication hardware may be qualified 

for use on the production line (where the features may be 

read and registered in a secure database). 

While more advanced authentication algorithms are being 

developed, it should be noted that this was not the purpose 

of this paper. The purpose was to use extremely simple 

authentication algorithms and inexpensive hardware for authentication, 

and demonstrate how high density security deterrents 

can be created through layering. The deployment 

recommendations are to use 1.251.25 mm 2 color tiles 

with an appropriate error-code checking (ECC) algorithm 

(e.g., Ref. 17), and to use 0.630.63 mm 2 binary tiles, also 

with an appropriate ECC algorithm. However, before deploying 

these security printing features, we would also perform 

a large set of qualification tests at and near the deployment 

size. This is necessary to predict more tightly the actual 

deployment error rate. Typically, one will perform many 

(hundreds or thousands) of tests at this more restricted 

range (e.g., at 1.1, 1.2, 1.3, and 1.4 mm dimensions for the 

color tiles), using multiple pieces of printing and scanning 

hardware. 

It is worth noting that feature density is not the only 

consideration in choosing between color and binary tiles. 

Color tiles provide a more difficult to reproduce look and 

feel, and may also “degrade” more quickly near the deployment 

tile size than binary tiles (as evidenced by a better 

“S”-shape in Fig. 6 when compared to Fig. 7). Moreover, 

color tiles can be “pretreated” for color space shifts that 

occur between the printing and scanning processes. For example, 

if the red and magenta tiles are found to be difficult 

to distinguish during authentication, additional blue may be 

added to the magenta and/or additional yellow may be 

added to the red. Additional color combinations can also be 

tested with the qualification protocol described here. In this 

way, color can be used to optimize the density of information 

encoded. 

The color tile features, in addition, are a means of fulfilling 

FDA recommendations for overt, covert, and forensic 



anticounterfeit technologies. 18 Clearly, the visible color patterns 

are an overt feature and can be used for branding in 

addition to product track and trace and authentication. The 

text encoded in the sequence of tiles provides a covert deterrent 

(visible, but not generally intelligible). The microtext 

superimposed on the color tiles, if deployed, can offer a 

forensic-level feature because the microtext fonts themselves 

can be varied with an astronomical number of 

combinations 8 that must be hand-authenticated. 

In addition to the color tile and binary tile 

qualification, 19 the use of multiple layers was considered. 

Because this “sandwich printing” feature is commercially 

available, it does not need qualification. On the HP Indigo 

digital press, sandwich printing is used for a variety of applications, 

one of which is a peel-off label. 10 Sandwich printing 

is possible due to this press’ ability to print as many as 16 

layers of ink on a substrate in a single pass (or “shot”) with 

perfect registration. The “sandwich” refers to the “front” 

design, the “back” design, and the opaque layer (the 

“cheese” of the sandwich, usually white ink) between them. 

When a transparent substrate is used for this layered design, 

there are two images created, each one visible from one side 

of the substrate. The opaque layer separates these two images. 

The layers of (usually white) ink between the ink layers 

for the two images serve two purposes: they provide the side 

which is currently viewed with a white underground and 

they hide the layer (against the substrate) that is behind. 

While the LEP ink (ElectroInk) is not opaque, it has roughly 

the transparency of an intentionally transparent screen 

printing ink. Thus, for it to block light between the two 

images in the layers of the sandwich, it must be applied in 

multiple layers. This is achieved through providing a separation 

in the print job for the opaque ink (usually white 

ink). Using sandwich printing, two layers of tiles, one for 

overt protection and the underlying second set for covert 

protection, can be layered, or “sandwiched,” over the same 

area on a package or document. This doubles the byte density 

possible for the layered deterrent. With sandwich printing, 

the layered deterrent described here can provide more 

than 3600 bits/in 2 , or 560 bits/cm 2 , of information. 

Thus, 1.8 cm 2 is required to provide 1024 bit security identifiers, 

which can be authenticated with inexpensive, commercially 

available scanners (without magnification). 


The authors gratefully acknowledge Jordi Arnabat, David 

Auter, Dan Briley, Maureen Brock, Carlos Martinez, Philippe 

Mücher, Andrew Page, Henry Sang, Eddie Torres, Juan Carlos 

Villa, and other colleagues for their assistance with aspects 

of this work. 

REFERENCES 

1 International Chamber of Commerce, Counterfeiting Intelligence 

Bureau, Countering Counterfeiting (ICC Publishing SA, Paris, France, 

1997). 

2 D. M. Hopkins, L. T. Kontnik, and M. T. Turnage, Counterfeiting 

Exposed (Wiley, Hoboken, NJ, 2003). 

3 K. Eban, Dangerous Doses (Harcourt, Orlando, FL, 2005). 

4 U.S. Food and Drug Administration, Medwatch, the FDA Safety 

Information and Adverse Event Reporting Program, website, http:// 

www.fda.gov/medwatch/. 

5 R. G. Johnston, “An anti-counterfeiting strategy using numeric 

tokens,“Int. J. Pharmaceutical Medicine (in press), also posted at: 

http://verifybrand.com/pdf/Drug_Anti-Counterfeiting_2004.pdf. 

6 S. J. Simske and R. Falcon, “Variable data security printing and the 

layered deterrent”, DigiFab 2005 (IS&T, Springfield, VA, 2005) pp. 

124–127. 

7 S. J. Simske and D. Auter, “A secure printing method to thwart 

counterfeiting”, HP Docket No. 200407401, filed with the US Patent and 

Trademark Office March 9, 2005. 

8 S. J. Simske, D. Auter, A. Page, and E. Torres, “A secure printing feature 

for document authentication”, HP Docket No. 200500190, filed with the 

US Patent and Trademark Office August 1, 2005. 

9 S. J. Simske, L. Ortiz, M. Mesarina, V. Deolalikar, C. Brignone, and G. 

Oget, “Ink coatings for identifying objects”, HP Docket No. 200405356, 

filed with the US Patent and Trademark Office October 12, 2004. 

10 S. J. Simske, P. Mücher, and C. Martinez, “Using variable data security 

printing to provide customized package protection”, Proc. IS&T’s 

DPP2005 (IS&T, Springfield, VA, 2005) pp. 112–113. 

11 Anoto substitute black ink, SunChemical AB, P.O. Box 70, 

Bromstensvagen 152, SE-163 91 SPANGA Sweden. 

12 S. J. Simske, “Low resolution photo/drawing classification: Metrics, 

method and archiving optimization”, Proc. ICIP 05 (IEEE, Piscataway, 

NJ, 2005). 

13 S. J. Simske, D. Li, and J. Aronoff, “A statistical method for binary 

classification of images”, DocEng 2005 (ACM, New York, NY, 2005) pp. 

127–129. 

14 F. M. Wahl, K. Y. Wong, and R. G. Casey, “Block segmentation and text 

extraction in mixed/image documents”, Comput. Vis. Graph. Image 

Process. 20, 375–390 (1982). 

15 J. Picard, C. Vielhauer, and N. Thorwirth, “Towards fraud-proof ID 

documents using multiple data hiding technologies and biometrics, 

“Proc. SPIE /ISSN 0-8194-5209-2, 416–427 (2004). 

16 Data Matrix, http://en.wikipedia.org/wiki/Data_Matrix. 

17 Reed-Solomon error correction, http://en.wikipedia.org/wiki/Reed- 

Solomon_error_correction. 

18 FDA Counterfeit Drug Task Force Interim Report, U.S. Department of 

Health and Human Services, FDA, also posted at: http://www.fda.gov/oc/ 

initiatives/counterfeit/report/interim_report.pdf, 46 pp., 2003. 

19 S. J. Simske, J. S. Aronoff, and J. Arnabat, “Qualification of security 

printing features”, Proc. SPIE in press. 




Preparation of Gold Nanoparticles in a Gelatin Layer Film 

Using Photographic Materials (5): Characteristics of Gold 

Nanoparticles Prepared on an Ultrafine Grain 

Photographic Emulsion 

Ken’ichi Kuge, Tomoaki Nakao, Seiji Saito, Ohiro Hikosaka and Akira Hasegawa 

Faculty of Engineering, Chiba University, Yayoi-cho, Inage-ku, Chiba 263-8522, Japan 

E-mail: kuge@faculty.chiba-u.jp 

Abstract. The authors report a process for the preparation of gold 

nanoparticles in a gelatin layer. This process is similar to the photographic 

process of gold development or gold latensification, where 

gold atoms are deposited on the exposed area of photographic material 

when it is immersed in a gold(I) thiocyanate complex solution. 

Gold particles have gained prominence for their nonlinear optical 

effect, the intensity of which depends on the density of the particles 

in the layer. The authors attempted to condense the particles using 

a photographic plate for hologram recording; this plate was made of 

an ultrafine grain emulsion because this emulsion was believed to 

be conducive to condensation. The characteristics of the particles 

were analyzed using photographic characteristic curves, absorption 

spectra, and size distributions. The characteristic curves rose gradually 

with the immersion period and finally showed a very high contrast 

curve. A sharp and strong plasmon absorption was observed at 

around 550 nm at high exposure values, while the peak exhibited a 

redshift and broadening at lower exposure values. The diameter of 

the particle increased proportionally with the square root of the immersion 

period. The growth rate decreased with the exposure value 

and was larger with high intensity exposure. The dependence on the 

exposure value was explained by the competition for the gold ion 

due to the high density of latent image specks. The larger growth 

rate with high intensity exposure was also explained by the low density 

of the latent image specks due to high intensity reciprocity 

failure. 



INTRODUCTION 

Gold particles have gained prominence for their nonlinear 

optical effects 1–4 and other useful characteristics. This optical 

effect will be useful for constructing nonlinear optical devices 

such as light switches or optical modulators. However, 

it is indispensable to prepare a film or solid state construction 

with dispersed gold particles in order to utilize gold 

particles in optical devices. We propose a new method to 

prepare gold particles dispersed in a gelatin layer by using 

photographic films, 5–10 wherein gold particles are prepared 

by immersing the exposed film in a gold(I) thiocyanate complex 

solution. Gold atoms are deposited on a latent image 

speck. By fixation, silver bromide grains are removed, and 

gold particles are left behind in the gelatin layer. This process 

is an application of the gold development process, 11–13 which 

Received Feb. 21, 2006; accepted for publication Jul. 6, 2006. 

1062-3701/2007/511/96/6/$20.00. 

produces an image of metallic gold, or gold latensification, 14 

wherein gold atoms are deposited on a latent image speck to 

achieve developability in silver halide photography. 

The preparation process of gold particles has attracted 

great interest and has been studied widely; however, knowledge 

regarding this process is still limited. Previously, we 

proposed that the deposition of gold atoms proceeded by the 

disproportionation reaction of three gold(I) ions to one 

gold(III) ion and two gold atoms catalyzed by latent image 

specks. 7 This process is similar to that of gold latensification 

proposed by Spencer et al. 15 Further, we reported on the 

reaction process, namely that the growth rate of gold atoms 

increased with the concentration of gold ions and that the 

diameter of gold particles increased proportionally with the 

square root of the immersion period. 8 

Meanwhile, Goertz et al. measured the Hyper-Rayleigh 

scattering (HRS) of AgBr nanosol decorated with sensitization 

centers. 16 The addition of KAuCl 4 enhanced the HRS of 

the nanosol with small silver clusters. They believed that this 

was due to the formation of Au atoms by the disproportionation 

reaction catalyzed by silver clusters followed by the 

incorporation of Au atoms into catalytic silver clusters. They 

found that the enhancement of HRS by the addition of 

KAuCl 4 reached saturation at higher densities of KAuCl 4 

and suggested that this process was self-limiting. 

The intensity of the nonlinear optical effect of gold particles 

depends on the density of the particles in a layer; 2,4 the 

higher the density, the stronger is the intensity. Therefore, 

condensation of the particles is necessary to enhance the 

nonlinear optical effect. Condensation is also important for 

utilizing gold particles in other applications. 17,18 

As one gold particle is formed on one latent image 

speck, the characteristics of the particle and the dispersing 

layer depend on how the latent image specks are prepared. 

Increasing the density of latent image specks in a layer is 

effective in increasing the density of gold particles. Previously, 

we proposed two possible methods to increase this 

density. 10 The first involves increasing the number of specks 

on a silver halide grain; this can be achieved by enhancing 

the dispersion of latent image specks. The second involves 

increasing the number of silver halide grains in a layer; this 

96

Kuge et al.: Preparation of gold nanoparticles in a gelatin layer film using photographic materials 5 

Figure 1. Photographic characteristic curves of gold particles in a gelatin 

layer at different immersion periods during gold deposition development. 

Left: high-intensity exposure for 10 −6 s; right: low-intensity exposure for 

100 s. 

Figure 2. Photographic characteristic curves for gold deposition development 

and normal development. Open circles and solid line: gold deposition 

development, 24 h, 20 °C; closed circles and dashed line: normal 

development using a D72 developer diluted to 1:4, 5 min, 20 °C. 

can be achieved by using an ultrafine grain emulsion. The 

result in our previous paper suggested that the latter technique 

was more effective. 10 

We then carried out observations for photographic materials 

that use an ultrafine grain emulsion, such as photographic 

plates for hologram recording, which are used for 

recording very fine diffraction gratings in the submicrometer 

range. We prepared gold particles by using a 

holographic plate and report the results of experiments using 

this plate. 

EXPERIMENT 

The photographic plate for hologram recording (P-5600, 

Konica-Minolta) was used as the sample photographic material. 

Ultrafine silver iodobromide grains with a diameter of 

35 nm were coated on a glass plate with high silver coverage, 

with the assumption that the coating was subjected to a 

rigorous hardening treatment in the production process. 

Two types of light source were used for exposure. One 

was a high-intensity (HI) xenon flash lamp with an exposure 

period of 10 −6 s, and the other was a low-intensity (LI) 

tungsten lamp with an exposure period of 100 s. The LI 

exposure was given using the JIS III sensitometer through a 

step tablet. The flash lamp for the HI exposure was set in the 

sensitometer, and exposure was given through the step tablet 

in this case as well. 

The formula of the gold complex solution for gold 

deposition was similar to that used in earlier studies. 7–10 The 

concentrations of the gold ion, potassium thiocyanate, and 

potassium bromide were 1.010 −3 , 1.210 −2 , and 

8.010 −3 mol/l, respectively. The exposed plates were immersed 

in the complex solution at 20 °C for 5–40 h.We 

also carried out normal development using a D72 developer 

diluted to 1:4. The development period and temperature 

were 5 min and 20 °C, respectively. Fixation was carried out 

with a normal F-5 photographic fixer for 5 min after the 

completion of gold deposition or normal development. 

We analyzed the characteristics of the gold particles using 

photographic characteristic curves, absorption spectra, 

and size distributions. The optical density (OD) of the plate 

with gold particles in the gelatin layer was measured with a 

densitometer through a green filter, and the characteristic 

curves for OD corresponding to green light were obtained. 

The absorption spectra of the same plate were measured 

with a double-beam spectrometer (Shimazu, UV-2600). The 

size distributions of the gold particles were obtained from 

observations with a transmission electron microscope 

(TEM) (JEOL 1200 Ex). We prepared the samples for TEM 

observation by applying the following suspension technique. 

The gelatin layer with the gold particles was scratched off 

from the plate and decomposed in an enzyme solution. The 

suspension with the gold particles was then dropped onto a 

grid covered with a collodion layer. 

EXPERIMENTAL RESULTS 

Photographic characteristic curves for gold deposition development 

for different immersion periods are shown in Fig. 1. 

The left part of the figure depicts the case of HI exposure for 

10 −6 s, while the right part depicts that of LI exposure for 

100 s. Since different light sources are used for the HI and LI 

exposures, a comparison between the exposure values with 

regard to the two intensities is irrelevant. The OD increased 

with the immersion period, and very high contrast curves 

were obtained for long immersion periods. 

Similar high contrast curves were also obtained in the 

case of normal development. The characteristic curves by 

gold deposition or normal development for the LI exposure 

are shown in Fig. 2. The sensitivity obtained by the former is 



Figure 3. Increasing rates of optical density at different exposure values. 


100 s. 

lower—about one-fifth that obtained by the latter. 

An increase in the OD of gold particles with the immersion 

period for the HI (left) and LI (right) exposures is 

shown in Fig. 3. The OD increased rapidly at higher exposure 

values for both intensities. Further, the OD at the HI 

exposure increased constantly, while the OD at the LI exposure 

reached saturation at longer immersion periods. 

The layer with the gold particles has a red-purple color 

at high exposure values and a blue or blue-purple color at 

low exposure values. The absorption spectra clearly exhibit 

these characteristics. Examples of spectra of the layer with 

different exposure values for the HI (left) and LI (right) 

exposures are shown in Fig. 4. The figures at the top correspond 

to the high exposure values, while those at the bottom 

correspond to low exposure values. The spectra at high exposure 

values show a sharp plasmon absorption by the gold 

particles peaked at around 550 nm. The peak absorbance 

increased with the immersion period, but the spectra continued 

to exhibit a sharp peak at the same wavelength. On 

the other hand, the spectra at low exposure values became 

broad, and the peak shifted to a longer wavelength with the 

immersion period. This can be attributed to a shift from 

plasmon to bulk absorption with an increase in the size of 

gold particles. We had observed the redshift and broadening 

of the peak with the immersion period in the previous results 

as well. 7 

The tendency shown in Fig. 4 is more pronounced in 

Fig. 5. The relationship between the peak wavelength and 

peak absorbance of the layer for the HI (left) and LI (right) 

exposures at different exposure values are shown in Fig. 5. 

An increase in the peak absorbance at the high exposure 

value caused only a small shift in the peak wavelength. On 

the other hand, a large redshift of the peak occurred at the 

low exposure value. 

The gold particles were observed with a TEM, and their 

diameters were measured using electron micrographs. A histogram 

of the diameter for each immersion period is shown 

in Fig. 6. The figure to the left shows the result for the HI 

exposure and a low exposure value, while the one to the 

Figure 4. Absorption spectra of gold particles in a gelatin layer for different 

immersion periods. Figures to the left: high-intensity exposure for 

10 −6 s; top: high exposure values, bottom: low exposure values. Figures 

to the right: low-intensity exposure for 100 s; top: high exposure values, 

bottom: low exposure values. Comparison of the exposure values between 

the figures to the left and those to the right is irrelevant as the light 

sources are different. 

Figure 5. Relationship between peak wavelength and peak absorbance 

at different exposure values. Left: high-intensity exposure for 10 −6 s; right: 

low-intensity exposure for 100 s. 

right shows that of the LI exposure and a high exposure 

value. The mean diameter increased with the immersion period, 

thereby broadening the distribution. The histograms 

that take into account other conditions indicated that the 

growth rate of the diameter was greater in cases with the 

lower exposure value and HI exposure. 

The growth rates at different exposure values for the HI 

(left) and LI (right) exposures are shown in Fig. 7. The 

curves were all convex to the upper site. The growth rate 



Figure 6. Size distribution of gold particles at different deposition periods. Left: high-intensity exposure for 

10 −6 s, with low exposure values of log rel.E=0.83; right: low-intensity exposure for 100 s, with high exposure 

values of log rel.E=1.18. 

corresponding to the HI exposure was greater than that corresponding 

to the LI exposure, and both the rates decreased 

with the exposure value. These figures differ from those corresponding 

to an increase in the rate of OD shown in Fig. 2, 

where OD increased monotonously and the rate of increase 

of OD was greater at the higher exposure value. 

Logarithmic plots of the diameter against the immersion 

period were straight lines, and the slopes of the lines 

were approximately equal to 0.5 for both the intensities and 

all exposure values. This suggests that the diameter d increases 

proportionally with the square root of the immersion 

period t, that is 

d = At 1/2 . 

The curves in Fig. 7 are the best fits to Eq. (1) and the 

experimental results were found to fit quite well to Eq. (1). 

The term A in Eq. (1) represents the rate constant, and a 

large value of A represents a large rate of increase. The relationship 

between the rate constant and the exposure value is 

shown in Fig. 8. The value of A decreased with the exposure 

value at both the intensities and was greater for the HI exposure 

than for the LI exposure, although we could not 

directly compare the exposure values as the light sources 

were different. Moreover, the value of A might reach saturation 

at a higher exposure value for both the intensities. 

The absorption spectra and the diameter histograms 

seem to suggest that the size distribution would be wider for 

lower exposure values. In order to verify this, the relationships 

between the diameter and the standard deviation are 

shown in Fig. 9. However, this figure reveals a result that 

contradicts the above expectation. The standard deviation 

increased with the diameter; however, it remained approximately 

constant regardless of the exposure value for the 

same diameter. The same tendency was observed at both the 

intensities, except that the standard deviations at the LI exposure 

were slightly greater than those at the HI exposure. 

Thus, for a particular mean diameter, the size distribution 

1 

Figure 7. Growth rates of particle diameter at different exposure values. 


100 s. 

does not change with the exposure values; thus, the size 

distribution itself does not depend on the exposure value. 

DISCUSSION 

As expected, the ultrafine grain emulsion produced fine gold 

nanoparticles dispersed in a gelatin layer with high density; 

this resulted in a sharp and strong plasmon absorption. 

Since the emulsion coating has a high density of grains, a 

considerably larger number of latent image specks are generated 

in the area corresponding to a high exposure value; 

this results in the high density of gold particles. Therefore, 

an ultrafine grain emulsion is suitable to obtain a high density 

of gold particles. 

The rapid increase in absorbance with the immersion 

period at high exposure values suggests an increase in the 

total number of gold atoms. This can be attributed to an 

increase in the size or number of the gold particles. However, 

the tendency of the gold particle to increase in size does 

not correlate with the increase in absorbance. The growth 

rate of the diameter at high exposure values was smaller than 

that at low exposure values, while the rate of increase of 

absorbance exhibited reverse characteristics. Therefore, an 



Figure 8. Relationship between exposure value and rate constant at different 

exposure values. Open circles and solid line: high-intensity exposure 

for 10 −6 s; closed circles and dashed line: low-intensity exposure 

100 s. 

increase in the number of gold particles must be the primary 

contributor to the increase in absorbance at high exposure 

values. This suggests that new particles should always be 

continuously formed over course of gold deposition. Unfortunately, 

we do not have sufficient knowledge of the rate of 

increase in the number of gold particles. On the other hand, 

the growth rate of the diameter was greater at lower exposure 

values. The absorption spectra simultaneously exhibited 

a redshift and broadening of the peak at lower exposure 

values; this suggested a shift from plasmon to bulk absorption 

of the larger gold particles. Consequently, an increase in 

both the number and diameter of the particles occurred at 

the low exposure values. 

Spatial distribution of particles in a layer also affects the 

optical characteristics. The effect of separation distance between 

silver particles on this optical characteristics has been 

discussed. 19 Similarly, a difference in exposure value should 

have some influence on the optical characteristics, as the 

distance of latent image specks should also vary with the 

exposure value. However, this is a rather complicated phenomenon, 

and the analysis will be reported in the future. 

The increase in diameter was proportional to the square 

root of the immersion period. A similar result had been 

obtained previously 7 as well and was discussed by Matejec 

and others 20,21 by citing an analysis of the growth rate in 

physical development. This analysis considered that the rate 

of increase in the number of silver particles m was proportional 

to the surface area of the particle S in the case of a 

reaction limited process 

dm/dt = k r S. 

The solution of this equation suggested that the diameter d 

of silver particles increased proportionally with the development 

period. 

On the other hand, the rate of increase in the number of 

silver particles was proportional to the diameter of the particle 

in the case of a diffusion limited process 

2 

Figure 9. Relationships between particle diameter and standard deviation 

at different exposure values. Left: high-intensity exposure for 10 −6 s; 

right: low-intensity exposure for 100 s. 

dm/dt = k d d. 

This indicated that the diameter increased proportionally 

with the square root of the development period. 

The analysis used for deriving Eq. (3) is based on the 

consideration that the chemical species pass through a thin 

diffusion layer around a latent image speck. This period of 

passage would be comparable to the normal development 

period of several minutes. On the other hand, gold deposition 

development required a longer period of several hours. 

It seems nearly impossible to regard the period of passage of 

the gold ions through the layer as several hours and thus 

simply apply the same analyses to the system of gold deposition. 

However, if the rate of increase in the number of the 

gold particles is proportionate to the diameter under certain 

conditions, a rate equation similar to Eq. (3) could be derived, 

which would lead to the growth rate given by Eq. (1). 

The exposure value significantly affects the formation 

process of gold particles since the growth rate was found to 

decrease with the exposure value. When distributions with 

the same immersion periods were compared, the particle size 

was found to be larger and the spread of the size distribution 

wider at lower exposure values. However, when we compared 

distributions with the same diameter, the standard deviation 

was approximately the same regardless of the exposure 

value, as shown in Fig. 9. This suggests that the size 

distribution remains the same at constant diameter and that 

the exposure value affects only the growth rate. The sharp 

and strong plasmon absorption at high exposure values is 

due to the large number of small gold particles growing 

slowly, while the redshift and broadening at low exposure 

values is due to the increase in the number of larger gold 

particles growing rapidly. 

The dependence on the exposure value is well explained 

insofar as the growth rate is affected by the supply of gold 

ions. The diffusion limited process is a case that meets this 

condition. At high exposure values, almost all grains have 

one or more latent image specks; therefore, the density of 

specks on the ultrafine grain emulsion is very high. As the 

absorbance increases continuously at high exposure values, 

new gold particles should be generated in succession on every 

latent image speck. In this case, many gold particles 

3 



compete to capture the gold ions; thus, the supply rate of 

gold ions should decrease, which might result in a low 

growth rate of gold particles. On the other hand, at low 

exposure values, the density is low. Therefore, sufficient 

number of gold ions can be supplied which would result in 

a larger growth rate. 

The result in Fig. 2, in which the sensitivities achieved 

by gold deposition and normal development are compared, 

indicated that a higher exposure value was necessary to trigger 

gold deposition. Some latent image specks did not trigger 

gold deposition; however, they triggered normal development. 

Therefore, the catalytic activity of latent image 

specks for gold deposition depends on their size, which is 

similar to the case of developability in normal development. 

The size of the gold particles for catalytic activity in gold 

deposition must be larger than that in normal development. 

The treatment for gold latensification is similar to that 

for gold deposition, except for the treatment period. In gold 

latensification, gold atoms are deposited even on the smaller 

silver specks of a latent subimage speck, thereby providing 

developability to these specks. However, they do not grow to 

gold particles by prolonged immersion. Only larger latent 

image specks can trigger continuous gold deposition. One 

possible explanation for this is that the catalytic activity of 

the silver atom is greater than that of the gold atom and that 

the incorporation of gold atoms into a silver atom cluster 

decreases the catalytic activity, although the mixture possesses 

normal developability. Based on HRS spectroscopy, 16 

Goertz et al. suggested that the incorporation process of gold 

atoms into a silver cluster by the same disproportionation 

reaction may be self-limiting. Therefore, a much larger size 

would be necessary for the gold and gold-silver clusters to 

exhibit catalytic activity on continuous immersion. 

Exposure intensity also seemed to affect the formation 

process of gold particles. The growth rate corresponding to 

the HI exposure was greater than that corresponding to the 

LI exposure. The explanation for this observation may be 

more complicated. One speculation is that high intensity 

reciprocity failure is significant, and latent image specks are 

not formed on every emulsion grain during the HI exposure. 

Some grains do not have latent image specks and this 

causes a decrease in the density of latent image specks, which 

in turn results in an increase in the growth rate. 

CONCLUSION 

We prepared gold particles using an ultrafine grain emulsion 

and successfully condensed them in a gelatin layer. We also 

analyzed the preparation process of gold particles. A sharp 

and strong plasmon absorption was observed at around 

550 nm at a high exposure value, while a redshift and broadening 

of the absorption due to a shift to bulk absorption of 

metallic gold appeared at a low exposure value. In addition, 

the growth rate of the particle diameter decreased with the 

exposure value. Therefore, new gold particles were generated 

in succession at high exposure values; this retarded the 

growth of the other gold particles. The high density of latent 

image specks at high exposure values resulted in a competition 

for gold ions, which brought about a decrease in the 

growth rate. The particle diameter increased proportionally 

with the square root of the immersion period. According to 

a previously reported analysis on physical development, this 

corresponds to the case wherein the rate-determining step 

was diffusion limited. It is speculated that the larger growth 

rate corresponding to the HI exposure is due to high intensity 

reciprocity failure, which caused a decrease in the density 

of latent image specks. 

REFERENCES 

1 D. Richard, P. Roussignol, and C. Flytzanis, Opt. Lett. 10, 511 (1985). 

2 F. Hache, D. Richard, and C. Flytzanis, J. Opt. Soc. Am. B 3, 1647 

(1986). 

3 F. Hache, D. Richard, C. Flytzanis, and U. Kreiberg, Appl. Phys. 47, 347 

(1988). 

4 H.Ma,P.Sheng,andG.K.L.Wong,Topics in Applied Physics 82, V.M. 

Shalaev, Ed. (Springer, Berlin, 2002) pp. 41–61. 

5 K. Kuge, M. Arisawa, N. Aoki, and A. Hasegawa, Jpn. J. Appl. Phys., Part 

1 39, 6550 (2000). 

6 K. Kuge, K. Kimijima, M. Arisawa, N. Aoki, and A. Hasegawa, J. Soc. 

Photogr. Sci. Technol. Jpn. 64, 242 (2001). 

7 K. Kuge, I. Suzuki, N. Aoki, and A. Hasegawa, J. Soc. Photogr. Sci. 

Technol. Jpn. 65, 536 (2002). 

8 T. Hirose, M. Arisawa, T. Omatsu, K. Kuge, A. Hasegawa, and M. Tateda, 

Proc. SPIE 4923, 127 (2002). 

9 K. Kuge, S. Saito, O. Hikosaka, T. Hirose, T. Omatsu, and A. Hasegawa, 

Particles 2003 (Toronto, Canada, 2003) p. 103. 

10 K. Kuge, M. Arisawa, N. Aoki, and A. Hasegawa, J. Imaging Sci. Technol. 

52, 176 (2004). 

11 T. H. James, J. Colloid Sci. 3, 447 (1948). 

12 D. C. Birch, G. C. Farnell, and R. B. Flint, J. Photogr. Sci. 23, 249 (1975). 

13 S. Jablonka, C. Mora, P. M. Nowak, and A. Zaleski, J. Inf. Rec. Mater. 23, 

249 (1996). 

14 T. Tani, Photographic Sensitivity (Oxford University Press, New York, 

1995) Chap. 6.3. 

15 A. C. Reeder and H. E. Spencer, J. Imaging Sci. 31, 126 (1987). 

16 M. Goertz, A. Rassbach, and M. R. V. Sahyun, J. Imaging Sci. Technol. 

46, 473 (2002). 

17 K. Kuge, Y. Goto, T. Nakao, O. Hikosaka, T. Sakai, and A. Hasegawa, J. 

Soc. Photogr. Sci. Technol. Jpn. 67, 490 (2004). 

18 K. Kuge, T. Nakao, Y. Goto, and A. Hasegawa, AgX2004: International 

Symposium on Silver Halide Technology (IS&T, Springfield, VA, 2004) pp. 

183–186. 

19 D. R. Whitcomb, S. Chen, J. D. Shore, P. J. Cowdery-Corvan, and K. A. 

Dunn, J. Imaging Sci. Technol. 49, 370 (2005). 

20 R. Matejec, Photogr. Korr. 104, 153 (1968). 

21 H. Kobayashi, T. Ohno, and S. Mizusawa, J. Soc. Photogr. Sci. Technol. 

Jpn. 44, 333 (1981). 


Journal of the Imaging Society of Japan VOL.45 NO.6 

2006 

CONTENTS 

Original Papers 

Mechanisms of Gloss Development with Matte-coated Paper in Electrophotography 

Y. KITANO, T. ENOMAE and A. ISOGAI ...5042 

Properties of Toner Charge and Toner Mass Amount on Developing Roller for Mono-Component 

Developing System A. SHIMADA, M. SAITO and T. MIYASAKA ...514 12 

Development of New Electron Transport Material with High Drift Mobility 

T. FUJIYAMA, K. SUGIMOTO and M. SEKIGUCHI ...521 19 

Development of New Polymerization Oil-less Full Color Toner 

H. NAKAJIMA, S. MOCHIZUKI, F. SASAKI, A. KOTSUGAI, Y. ASAHINA, S. MATSUOKA, 

O. UCHINOKURA, S. NAKAYAMA, M. ISHIKAWA and K. SAKATA ...526 24 

Design of Coated Paper and Fading Characteristics of Dyes Using a New Rewritable System 

Y. HASHIMOTO, T. YUASA, N. MIYAMACHI, Y. NAITO, T. ISHIYAMA, 

S. NISHIDA, T. ASANO and D. TSUCHIYA ...532 30 

Imaging Today 

“Recent Technologies of Color Printers Introduced from 2005 to 2006” 

Introduction K. MARUYAMA, T. TAKEUCHI and M. KIMURA ...540 38 

Technology Differentiation of CLP-300 Color Printer for Personal & SOHO 

M.-H. CHOI, K.-H. KIM, K.-J. PAK, S.-D. AN and Y.-G. KIM ...541 39 

Development ofFull-colorMFPDP-C322Series 

M. KAMATA, T. OZAKI, N. TAJIMA , M. SAMEI and K. TERAO ...546 44 

MICROLINE 9600PSMICROLINE Pro 9800PS Series 

N. OISHI, T. ASABA, Y. MURATA, M. YAJI and C. KOMORI ...553 51 

Development of Full Color MFP ApeosPort-II/DocuCente-II C4300 Series Y. YAKABE ...559 57 

Midrange to High-end Digital Full-color Multifunctional Printers imagio MP C3500 Series 

K. MATSUMOTO,Y. TAKAHASHI, A. AMITA and S. UENO ...567 65 

iRC-5180 SeriesEquippedwithNewFuserSystemTBFM. JINZAIY. KAMIYA and K. AOKI ...573 71 

Coated Paper Handling Technologies of the KONICA MINOLTA bizhub Pro C6500Y. ICHIHARA ...579 77 

Lectures in Science 

Introduction to Modeling and Numerical Simulation of Electrophotography (II) 

Finite Difference MethodH. KAWAMOTO and M. KADONAGA ...586 84 

Meeting Reports 593 91 

Announcements 597 95 

Guide for Authors 601 99 

Contents of J. Photographic Society of Japan 602100 

Contents of J. Printing Science and Technology of Japan603101 

Contents of J. Inst. Image Electronics Engineers of Japan 604102 

Contents of Journal of Imaging Science and Technology 606104 

Essays on Imaging 

The Imaging Society of Japan 

c/o Tokyo Polytechnic University, 2-9-5, Honcho, Nakano-ku, Tokyo, 1648768 Japan 

Phone :033373-9576 Fax :033372-4414 E-mail :

Advanced Measurement Systems for All R&D and 

Quality Control Needs in Electrophotography, 

Inkjet and Other Printing Technologies 

PDT ® -2000 series 

Electrophotographic characterization, 

uniformity mapping, and defect 

detection for large and small format 

OPC drums 

PDT ® -1000L 

PDT ® -1000 

ECT-100 TM 

OPC drum coating thickness 

gauge 

Electrophotographic 

Component Testing 

MFA-2000 TM 

Magnetic field distribution 

analysis in mag roller magnets 

DRA-2000 TM 

Semi-insulating components testing 

including charge rollers, mag rollers, 

transfer rollers, transfer belts, and 

print media 

TFS-1000 TM 

Toner fusing latitude testing 

Objective Print Quality 

Analysis for All Digital 

Printing Technologies 

IAS ® -1000 

Fully-automated high volume print 

quality testing 

Scanner-based high speed print 

quality analysis 

Scanner IAS ® 

Personal IAS ® 

Handheld series for print quality, 

distinctness of image (DOI), and 

color measurements. Truly portable; 

no PC connection required 

PocketSpec TM 

DIAS TM 

Quality Engineering Associates, Inc. 

99 South Bedford Street #4, Burlington, MA 01803 USA 

Tel: +1 (781) 221-0080 • Fax: +1 (781) 221-7107 • info@qea.com • www.qea.com

imaging.org 

your source for imaging technology conferences 

IS&T ANNOUNCES A NEW MEETING 

International Symposium 

on 

Technologies 

for 

Digital 

Fulfillment 

March 3-5, 2007 

Las Vegas, Nevada 

Register NOW for this exciting inaugural conference!!! 

• 26 Talks 

• 2 Working Demonstrations of New Technology 

Advances 

• Luncheon Keynote by Mark Schneider, CTO of 

Kodak’s Consumer Digital Group who will review 

special technological developments in the field of 

digital fulfillment 

• Conference Reception 

Technical Areas 

• Retail (kiosks or minilabs), Home (snapshot 

printers), Mobile (cell phone cameras and 

on-line photofinishing) 

• Output Media (AgX, thermal, EP, and Ink Jet) 

• Capture Devices (DSCs, film, camera phones) 

• Output Equipment (printers, minilabs, kiosks) 

• Image processing (workflow, algorithms) 

• Multiple Opportunities for Networking 

The first International Symposium on Technologies for Digital Fulfillment is an evolution of the highly-acclaimed biennial International Symposia 

on Photofinishing Technology. The 2007 conference will be held March 3-5, 2007, just prior to DIMA/PMA2007, 

at the Westin Casuarina Hotel and Spa in Las Vegas, Nevada. 

Details and registration information can be found at www.imaging.org/conferences/tdf2007 

Hotel reservations can be made at www.starwoodmeeting.com/Book/sist.

JIST - Society for Imaging Science and Technology

Create successful ePaper yourself

Delete template?

Save as template?