Printing Technique Classification for Document Counterfeit Detection

More documents

Recommendations

Info

(a)(b)Figure 1. Sample letter from different inkjetprintouts (top) and laser printouts (bottom).Laser printed letter typically show sharpercontours than inkjet printed ones.2.1 Per-Letter ClassificationA canonical choice for getting a hint on the genuinenessof a printed document is to study its creation process, inparticular the type of printer that has been used. However,little research has been done in this area so far.One exception are studies on the dot patterns of inkjetprinters by Yamashita et al.[3]. Tchan analyzed the sharpnessof image features to detect manipulations of existingprintouts[4]. A more specialized solutions is proposed e. g.by Akao et al., who search for spur marks, which inkjetprinters commonly leave on the paper during the printingprocess[5].The approach proposed in this paper is to generically detectforgeries by using a per-letter classification of the printingtechnique. It relies on the fact that for different printershave different visual characteristics, especially in the edgearea of the letters. From the grayscale distribution in thisimage area, it can be decided for each letter in the documentwhat kind of printer created it. In parts, this is similarto Tchan[6], but there the target was to distinguish differentprinter models, which required details knowledge on thecreation of the original document, access to the machines tobe detected, and also specialized hardware for the imageacquisition. The method we propose is more generic, it ispossible to target a wide range of different documents types.Another advantage is, that the method does not require anyspecific hardware but can work with a standard consumerimaging device like a scanner or a digital camera.The suitability of the proposed technique to counterfeitdetection is backed up by forensic results on how documentforgeries are usually done these days: only a very smallfraction of forgers are professionally equipped with an expensivelaboratory of advanced printing machinery. Instead,the vast majority of forgeries are done on an ordinary homePC with a scanner and a printer. This includes the replacementof passports photographs by inkjet printouts, and scanningbus-tickets and printing copies on a color laser printer.A surprisingly large number of forgeries is even simplyFigure 2. Enlarged contour region of a inkjetprinted (a) and a laser printed character (b).The laser image shows a sharper black-towhitetransition and fewer ink/toner dropletsoutside of the black contour.done by hand, e. g. adding or manipulating digits and aninvoice. It is characteristic for these kinds of forgeries, thatthe forged information relies on a different printing techniquethan the original. This can be the case that the wholedocument was created using a wrong technique, like for thebus-ticket, or that some parts have different, like on the invoice.Both scenarios can be detected by the proposed method:if the original printing technique of a document is known, itcan as a whole be validated against the detection techniquein the document to be checked. Additionally, documentswhich show only few positions of different printing techniquethan the main body show strong evidence of a manipulation.3 Printing Technique ClassificationThe proposed method consists of four steps: preprocessing,feature extraction, classification and visualization,which we describe in the following in more detail.3.1 Image PreprocessingFor a more compact description, we leave out the moreparts of image acquisition and assume that the system isprovided with high resolution image material. Since we aredealing with document images, we usually cannot rely oncolor information, and we therefore assume the material tobe in 8 bit grayscale format. We identify individual objectson the page by a connected component analysis on a binarizedversion of the image. Of the connected componentswe keep all which have roughly the right size and aspectratio to be characters and extract the classification featuresfrom them.3.2 Feature ExtractionWhen studying high resolution scans of different printertypes, sharp edges are the most interesting regions to distinguishbetween them visually. Here, the techniques differ,
(a) Letter Contours for Inkjet(b) Letter Contours for LaserFigure 3. Line Edge Roughness for Inkjet (a)and Laser (b) print. In low resolution, laserand inkjet letter contours are roughly equallysmooth (small letters). In full resolution, theinkjet print has rougher contours than thelaser one (large letters). Therefore, the ratiobetween high res and low res contour lengthscan be used as a distinguishing feature.e. g. in sharpness and existence of droplets. This is illustratedin Figure 2. From these areas, we extract the followingqualities to form a feature vector: Line Edge Roughness,Area Difference, Correlation Coefficient, Texture.Line Edge Roughness It is a basic observation that in veryhigh resolution images, the contour line of an inkjet printtypically is rougher than that of a laser print, and it thereforehas a longer perimeter. However, if the image is scaleddown, the contours of inkjet and laser printouts both becomesmoother and it is generally not possible to distinguishthem anymore. Figure 3 shows this effect.For classification, we calculate the ratio of the lengths ofthe high res and the low res contours, and we include thescale as a factor as well, because the roughness measurethen becomes invariant to the actual resolutions used:perimeter of downsized object contour ∗ scaleroughness =perimeter of object contour(1)For our experiments, we use a scale factor of 10.Area Difference Another characteristic property of inkjetprints is that the grayscale distribution of a black-to-whiteedge transition contains more intermediate grayscale valuesfor inkjet than for laser prints. To measure this in a sizeinvariantmatter, we use the area difference between twodifferent image binarizations. One is created using Otsu’sautomatic threshold [7], the other by using a threshold δunits larger. From the difference between the area A otsu ofthe first binary shape and the area A otsu+δ of the second binaryshape we obtain the number of intermediate grayscaleFigure 4. Area Difference: The size of a binarizedletter (left) is increased slightly by raisingthe binarization threshold (center). Thearea of the additional pixels (right) typicallyis larger for inkjet printers than for laser andcan be used for classification.pixels, which we normalize by the total area:area difference = |A otsu − A otsu+δ |A otsu(2)This is illustrated in Figure 4 using δ = 20, which is thesame value that we used in our experiments.Correlation Coefficient It is also characteristic for a printerhow close its output of a letter contour comes to an ideastep edge. We measure this by calculating the correlationcoefficient between the original image and the segmentedbinary image. Since only the contour region is importantto us, we use an edge image that had been dilated with acircular mask of radius 7 as a region-of-interest (ROI) mask.Figure 5 illustrates this.When A is the original gray value image and B the Otsubinarizedversion, the correlation coefficient is calculated as∑(A[i, j]−Ā)(B[i, j]− ¯B)[i,j]∈ROIcorrelation = √ ∑√ ∑(A[i, j]−Ā)2 (B[i, j]− ¯B) 2[i,j]∈[i,j]∈ROIROI(3)where Ā and ¯B are the mean of A and B respectively overthe ROI.Texture To obtain a texture descriptor, we use a techniquethat is similar to the gray value co-occurrence matrix, butmeasures the co-occurrence of values in two different imagesor channels. One is the original image, another is atransformed version of the original. We do this for threedifferent transforms:Gaussian Filter We obtain the the second channel fromblurring the original using a Gaussian filter.Wavelet Filter We obtain the second channel from computingthe first low pass approximation in a waveletsrepresentation of the original image.
Page 4 and 5: For each of the transformed images,
Page 6: (a) (b) (c)Figure 7. Example of cla

Printing Technique Classification for Document Counterfeit Detection

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?