Aditya kumar et al, Int. J. Comp. Tech. Appl., Vol 2 - International ...
Aditya kumar et al, Int. J. Comp. Tech. Appl., Vol 2 - International ...
Aditya kumar et al, Int. J. Comp. Tech. Appl., Vol 2 - International ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
<strong>Aditya</strong> <strong>kumar</strong> <strong>et</strong> <strong>al</strong>, <strong>Int</strong>. J. <strong>Comp</strong>. <strong>Tech</strong>. <strong>Appl</strong>., <strong>Vol</strong> 2 (4), 1099-1117<br />
ISSN:2229-6093<br />
Image <strong>Comp</strong>ression by Moment Preserving<br />
Algorithms: A Scrutinization<br />
<strong>Aditya</strong> Kumar<br />
Department Of <strong>Comp</strong>uter Science & Engineering<br />
Nation<strong>al</strong> Institute Of <strong>Tech</strong>nology, Hamirpur, India<br />
adi.bond.adi@gmail.com<br />
Pardeep Singh<br />
Department Of <strong>Comp</strong>uter Science & Engineering<br />
Nation<strong>al</strong> Institute Of <strong>Tech</strong>nology, Hamirpur, India<br />
pardeep@nitham.ac.in<br />
Abstract-Block Truncation Coding is an image compression<br />
technique which is used glob<strong>al</strong>ly in many online as well as offline<br />
graphic<strong>al</strong> applications e.g. LCD overdrive <strong>et</strong>c. In this literature<br />
survey we will discuss many variants of Block Truncation coding<br />
and latest m<strong>et</strong>amorphosed techniques which are introduced<br />
recently. These techniques are an<strong>al</strong>yzed beh<strong>al</strong>f on objective and<br />
subjective point of view. In this literature survey, simulation<br />
results are <strong>al</strong>so provided for each technique.<br />
Keywords- Economic<strong>al</strong> Block Truncation Coding, filtering,<br />
interpolation., LCD overdrive, transformation.<br />
I. INTRODUCTION<br />
Block Truncation Coding (BTC) [1] [2] [3] is a<br />
straightforward and accelerated lossy image compression<br />
technique for gray sc<strong>al</strong>e images invented by Delp and Mitchell<br />
[20]. The main conception of BTC is to execute moment<br />
preserving quantization for each blocks of each pixels by<br />
which the qu<strong>al</strong>ity of the image will be remained endorse and at<br />
the same time the request for the storage space will be<br />
reduced. Even if the compression using Block Truncation<br />
Coding is inferior to the standard JPEG compression <strong>al</strong>gorithm<br />
[21], BTC has achieved fame due to its experiment<strong>al</strong><br />
usefulness. Sever<strong>al</strong> variants related to the basic Block<br />
Truncation Coding have been discussed in this literature<br />
survey.<br />
L<strong>et</strong> us weigh up a digit<strong>al</strong> gray-level immobile image of the<br />
size N pixels where every single pixel is represented by k bits.<br />
The motto of the image compression is to convert or transform<br />
the image to a space efficient (compressed) form so that the<br />
data of image is protect as far as possible when decompressing<br />
the encoded image.<br />
The goodness of compression techniques can be<br />
characterized by considering a s<strong>et</strong> of qu<strong>al</strong>ity. These qu<strong>al</strong>ities<br />
are included the bit rate which gives the average number of<br />
bits per stored pixel of the image. Bit rate is the excellent<br />
limiting factor [4] [5] [23] of a compression technique because<br />
it measures the efficiency of the technique.<br />
Another goodness of this m<strong>et</strong>hod is the ability to safeguard<br />
the information data. A compression technique is lossless if<br />
the decompressed image is exactly the same as the origin<strong>al</strong><br />
one. Otherwise the inherit m<strong>et</strong>hod is lossy. The excellence of<br />
the reconstructed image can be measured by the mean square<br />
error (MSE) [6] [7], mean absolute error (MAE) [6] [7], sign<strong>al</strong><br />
to noise ratio (SNR) [1] [9], or it can be an<strong>al</strong>yzed by a human<br />
photo an<strong>al</strong>yst. L<strong>et</strong> x i stand for the i th pixel v<strong>al</strong>ue in the origin<strong>al</strong><br />
image and y i is the corresponding pixel in the reconstructed<br />
image (i=1,2,...,N). Here the row-major order of pixels of the<br />
image matrix is considered. The three quantitative measures<br />
are defined as<br />
MSE = ∑<br />
|y − x <br />
| <br />
(1)<br />
MAE = ∑ <br />
|y − x | (2)<br />
SNR = −10log <br />
(3)<br />
These above limiting factor are glob<strong>al</strong>ly used but<br />
c<strong>al</strong>amitously they do not <strong>al</strong>ways occur simultaneously with the<br />
estimations of human expert [23] [27].<br />
The third one characteristics of BTC is the processing<br />
speed of the compression and decompression <strong>al</strong>gorithms. In<br />
on-line applications the response times are often serious<br />
param<strong>et</strong>er. In the extreme case a space productive<br />
compression <strong>al</strong>gorithm is useless if its processing time causes<br />
an insufferable delay in the image processing related<br />
application. If image compression task work as a background<br />
process then it is tolerable. However, fast decompression is<br />
desirable [64].<br />
The fourth one characteristics of BTC is robustness against<br />
data transmission errors. The compressed file is norm<strong>al</strong>ly an<br />
object of a data transmission operation. The transmission is in<br />
the elementary form b<strong>et</strong>ween primary storage and secondary<br />
memory but it can be b<strong>et</strong>ween two remote sites via<br />
transmission lines. Visu<strong>al</strong> comparison of the origin<strong>al</strong> and<br />
disturbed images gives a direct ev<strong>al</strong>uation of the robustness.<br />
However, the data transmission systems commonly contain<br />
fault tolerant intern<strong>al</strong> data formats so that this property is not<br />
<strong>al</strong>ways effective or v<strong>al</strong>id.<br />
Another important features of the compression m<strong>et</strong>hod we<br />
may mention the ease of implementation and the demand for a<br />
working storage space. Present time these measurement<br />
param<strong>et</strong>ers are often of secondary importance. From the<br />
experiment<strong>al</strong> point of view the last but often not least feature<br />
is running time of the <strong>al</strong>gorithm itself. Certainty and<br />
IJCTA | JULY-AUGUST 2011<br />
Available online@www.ijcta.com<br />
1099
<strong>Aditya</strong> <strong>kumar</strong> <strong>et</strong> <strong>al</strong>, <strong>Int</strong>. J. <strong>Comp</strong>. <strong>Tech</strong>. <strong>Appl</strong>., <strong>Vol</strong> 2 (4), 1099-1117<br />
ISSN:2229-6093<br />
dependability of the software often to a high degree depends<br />
on the complexity of the <strong>al</strong>gorithm.<br />
After this introduction we have given BTC <strong>al</strong>gorithm in<br />
section II, variants of basic m<strong>et</strong>hods in section III,<br />
comparative an<strong>al</strong>ysis on recent BTC <strong>al</strong>gorithm in section IV,<br />
comparison to JPEG in section V and fin<strong>al</strong>ly the conclusion in<br />
last section VI.<br />
II.<br />
BTC ALGORITHM<br />
An image having N pixel is divided into non overlapping<br />
sm<strong>al</strong>ler blocks of the size m pixels and each individu<strong>al</strong> block<br />
is processed separately [64]. A condensed representation is<br />
forage through loc<strong>al</strong>ly for each block. While the decoding<br />
process m<strong>et</strong>amorphoses the encoded blocks back into pixel<br />
v<strong>al</strong>ues so that the decompressed image look like the origin<strong>al</strong><br />
one as much as possible. The mean v<strong>al</strong>ue (x̅) and standard<br />
deviation (σ) are enumerated and encoded for each block.<br />
x̅ = ∑<br />
= ∑<br />
x <br />
<br />
x <br />
(4)<br />
<br />
x <br />
(5)<br />
σ = x − x̅<br />
<br />
(6)<br />
Then a two-level quantization is electrocuted. Pixels with<br />
v<strong>al</strong>ues less than the quantization threshold ( x ≤ x ) are<br />
mapped to v<strong>al</strong>ue a and pixels with v<strong>al</strong>ues greater than or equ<strong>al</strong><br />
to the threshold (x > x ) are mapped to v<strong>al</strong>ue b. Here the<br />
threshold x th is considered to be x and the v<strong>al</strong>ue a and b are<br />
chosen so that the first and second sample moments (x̅, x ) are<br />
safe guarded in the encoding processes. This is the case for the<br />
selection<br />
a = x̅ − σ <br />
<br />
b = x̅ + σ <br />
<br />
where, q stands for the number of pixels x i >x th .<br />
A compressed block emerges as a triple (x̅, σ, B), where x̅<br />
and σ give mean and standard deviation of the pixel v<strong>al</strong>ues in<br />
the block and B stands for the bit plane, giving the<br />
quantization of the pixel v<strong>al</strong>ues [64] . See Figure1.<br />
A direct encoding of x̅ and σ by k bits yields a bit rate of<br />
<br />
<br />
(7)<br />
(8)<br />
= 1 +<br />
<br />
bits per pixel (bpp) (9)<br />
<br />
The m<strong>et</strong>hod is accelerated, to be deficient in extra space, is<br />
easy to carry out [23] [64], and has low computation<strong>al</strong><br />
demands. It conserved the qu<strong>al</strong>ity of the reconstructed image<br />
and r<strong>et</strong>ains the edges. It <strong>al</strong>so recovers excellently from<br />
transmission errors. One can still improve it in sever<strong>al</strong><br />
different ways.<br />
III.<br />
VARIANTS OF BASIC METHOD<br />
Many techniques are related to basic Block Truncation<br />
Coding have been discussed in this section. They focus<br />
glob<strong>al</strong>ly at a lower bit rate but other demerits of the m<strong>et</strong>hod<br />
are <strong>al</strong>so examined. The greatest deficiency of BTC is a<br />
relatively high bit rate as compared to other image<br />
compression m<strong>et</strong>hods, like DCT [9] [19], vector quantization<br />
[28] [31]. Other main demerits are ragged edges (staircase<br />
effect) regenerate in statue (image).<br />
Block Truncation Coding is divided into three separate<br />
steps, and each of these is an<strong>al</strong>yzed separately. The over<strong>al</strong>l<br />
structure of the <strong>al</strong>gorithm is as follows:<br />
BTC <strong>al</strong>gorithm for coding a single block<br />
1) Perform quantization<br />
- Select the threshold.<br />
- Select the quantization levels (a, b).<br />
2) Code the quantization data.<br />
3) Code the bit plane.<br />
In many variants, the choice of the threshold and the pair<br />
(a, b) are non dependable to each other, while some variants<br />
concern only one or the other. Most of the variants do not<br />
suppose a specific block size. Their instrumentations,<br />
however, are often planed for 4*4 blocks only.<br />
Origin<strong>al</strong> Bit Plane Reconstructed<br />
2 9 12 15 0 1 1 1 2 12 12 12<br />
2 11 11 9 0 1 1 1 2 12 12 12<br />
2 3 12 15 0 0 1 1 2 2 12 12<br />
3 3 4 14 0 0 0 1 2 2 2 12<br />
x=7.94<br />
σ=4.91<br />
q=9 a=2.3<br />
b=12.3<br />
Figure1. Block Truncation Coding by tripl<strong>et</strong> (x̅, σ, B) [64]<br />
A. Performing the Quantization<br />
In commencement, choosing the threshold pixel v<strong>al</strong>ue and<br />
choosing the quantization levels can be seen as two non<br />
corresponding stages. While most BTC techniques are<br />
encoded as safeguard certain sample moments, or to minimize<br />
some loy<strong>al</strong>ty function. Therefore x th , a and b are in close<br />
documentation with each other [64].<br />
Next part of this section will be included sever<strong>al</strong> variants,<br />
which can be characterized into three categories. The first<br />
category consists of the m<strong>et</strong>hods that will safeguard certain<br />
moments. The m<strong>et</strong>hods of the second category minimize the<br />
MSE-v<strong>al</strong>ue, and the m<strong>et</strong>hods in the third one minimize the<br />
MAE-v<strong>al</strong>ue. The remainders of the variants are intermediate<br />
b<strong>et</strong>ween these three types.<br />
a) Moment Preserving Quantization<br />
A most number of the changes found in this portion are<br />
based on the origin<strong>al</strong>, two moments preserving quantization<br />
[20] [24] [25]. The following formulas must hold for this:<br />
mx̅ = (m − q). a + q. b (10)<br />
IJCTA | JULY-AUGUST 2011<br />
Available online@www.ijcta.com<br />
1100
<strong>Aditya</strong> <strong>kumar</strong> <strong>et</strong> <strong>al</strong>, <strong>Int</strong>. J. <strong>Comp</strong>. <strong>Tech</strong>. <strong>Appl</strong>., <strong>Vol</strong> 2 (4), 1099-1117<br />
ISSN:2229-6093<br />
mx = (m − q). a + q. b (11)<br />
This will be grasped by using x̅ as a threshold and<br />
choosing a and b according to equation 7. This option is not<br />
have optim<strong>al</strong> v<strong>al</strong>ue according to MSE point of view. However,<br />
the base conception behind the technique has been a different<br />
loy<strong>al</strong>ty measurement for example safeguarding the sample<br />
moments of the input [27]. A motivation for the m<strong>et</strong>hod is that<br />
the human eye does not perceive sm<strong>al</strong>l changes of intensity<br />
b<strong>et</strong>ween individu<strong>al</strong> pixels, but is sensitive to the average<br />
(mean) v<strong>al</strong>ue and contrast in larger regions [27] [64].<br />
b) Third Moment Preserving Quantization<br />
One can <strong>al</strong>ter BTC so that it safe guard in addition to x<br />
and x the third moment x [20] [64], too. This can be gained<br />
by choosing x th, so that the param<strong>et</strong>er q attains the v<strong>al</strong>ue<br />
where,<br />
q = . 1 + A <br />
<br />
A =<br />
(12)<br />
̅ ̅<br />
(13)<br />
<br />
In distinction to the basic BTC the encoding phase is more<br />
necessitated. On the other hand, decoding is still fast. The<br />
point of the variant is that it upgraded some features (near<br />
edges) of the image. In fact, the third moment corresponds to<br />
the skewness of the distribution [23] [26]. Negative v<strong>al</strong>ues<br />
indicate negative skewness and positive v<strong>al</strong>ues positive<br />
skewness. Quantization levels (a,b) must be selected<br />
according to equation 7.<br />
c) Gener<strong>al</strong>ized Moment Preserving Quantization<br />
The BTC technique can be gener<strong>al</strong>ized to safe guard the r,<br />
2r and 3r th moments [15]. Here is the exact formulation of the<br />
m<strong>et</strong>hod. Higher moments cause overflow problems and rarely<br />
good qu<strong>al</strong>ity images. If one wants to preserve r and 2r th<br />
moments only, the threshold is<br />
x <br />
(14)<br />
<br />
<strong>Comp</strong>ared to the origin<strong>al</strong> first and second moments<br />
preserving quantization, the m<strong>et</strong>hod imperceptibly upgrades<br />
the MSE-v<strong>al</strong>ue [21].<br />
d) Absolute Moment BTC<br />
Absolute moment block truncation coding (AMBTC) by<br />
Lema and Mitchell [23] [24] [25] preserves x̅ and the sample<br />
first absolute centr<strong>al</strong> moment<br />
α = ∑ <br />
|x<br />
− x̅|<br />
(15)<br />
Quantization levels are c<strong>al</strong>culated from<br />
a = x̅ − <br />
. <br />
<br />
b = x̅ + <br />
. <br />
<br />
(16)<br />
(17)<br />
One can show that a=x (lower mean) and b=x (higher<br />
mean), where<br />
x =<br />
<br />
<br />
̅ (20)<br />
∗ ∑ |x |<br />
x <br />
= ∗ ∑ |x |<br />
<br />
̅ (19)<br />
After <strong>al</strong>l, this selection decrease the MSE [22] v<strong>al</strong>ue<br />
among the BTC-variants that use x as a quantization threshold<br />
[23]. It can, however, efficiently be shown in Table I and<br />
Table II that it is optim<strong>al</strong> for other selections of x too. The<br />
encoding and decoding tasks are very fast for AMBTC<br />
because square root and multiplication operations are omitted.<br />
e) MSE-Optim<strong>al</strong> Quantization<br />
The MSE-optim<strong>al</strong> choice for (a,b) is according to<br />
AMBTC, and it is non dependable of the quantization<br />
threshold. Therefore the MSE-optim<strong>al</strong> quantization can be<br />
obtained by choosing the quantization threshold so that it<br />
minimizes:<br />
<br />
MSE = ∑ x − a + ∑<br />
x − b (20)<br />
The MSE-optim<strong>al</strong> quantization is <strong>al</strong>so known as minimum<br />
MSE quantization (MMSE).<br />
f) MAE-Optim<strong>al</strong> Quantization<br />
The MAE loy<strong>al</strong>ity criterion (2) can be minimized by choosing<br />
a=median x 1 , x 2 ,…………………….,x m-q-1 (21)<br />
b=median x m-q , …………………….,x m (22)<br />
Now the optim<strong>al</strong> threshold is again found by an exhaustive<br />
search [20]. The MAE-optim<strong>al</strong> quantization is <strong>al</strong>so known as<br />
minimum MAE quantization (MMAE).<br />
g) HEURISTIC CRITERION<br />
Goeddel and Bass [25] proposed a heuristic selection<br />
criterion for the threshold:<br />
x = <br />
<br />
(23)<br />
Where x min and x max stand for the minim<strong>al</strong> and maxim<strong>al</strong><br />
pixel v<strong>al</strong>ues of a block. The criterion gives improved MSEv<strong>al</strong>ues.<br />
The moments will not be safe guarded by the<br />
technique. However, if we use formulas (16-17) to select (a, b)<br />
instead of (7), this criterion can be considered as a practic<strong>al</strong><br />
estimation for the MSE-optim<strong>al</strong> quantization.<br />
Table I<br />
GB+mean Lyod MSE-opt. MAE-opt<br />
x th (x max+x min)/2 (a+b)/2 MSE-opt. MAE-opt<br />
(a,b) Mean v<strong>al</strong>ues Mean v<strong>al</strong>ues Mean v<strong>al</strong>ues median<br />
MSE 37.05 36.83 35.54 40.21<br />
MAE 3.68 3.78 3.54 3.34<br />
SNR 32.44 32.47 32.62 32.09<br />
Table II<br />
GB+mean Lyod MSE-opt. MAE-opt<br />
x th (x max+x min)/2 (a+b)/2 MSE-opt. MAE-opt<br />
(a,b) Mean v<strong>al</strong>ues Mean v<strong>al</strong>ues Mean v<strong>al</strong>ues median<br />
MSE 37.05 36.83 35.54 40.21<br />
MAE 3.68 3.78 3.54 3.34<br />
SNR 32.44 32.47 32.62 32.09<br />
h) Lloyd Quantization<br />
Another closely MSE-optim<strong>al</strong> quantization is obtained by<br />
an iterative <strong>al</strong>gorithm proposed by Efrati and Lu [24]:<br />
IJCTA | JULY-AUGUST 2011<br />
Available online@www.ijcta.com<br />
1101
<strong>Aditya</strong> <strong>kumar</strong> <strong>et</strong> <strong>al</strong>, <strong>Int</strong>. J. <strong>Comp</strong>. <strong>Tech</strong>. <strong>Appl</strong>., <strong>Vol</strong> 2 (4), 1099-1117<br />
ISSN:2229-6093<br />
1. L<strong>et</strong> x th = x.<br />
2. <strong>Comp</strong>ute a and b according to (16) and (17).<br />
3. If x th = (a+b)/2 then exit.<br />
4. L<strong>et</strong> x th = (a+b)/2. Go to step 2.<br />
The <strong>al</strong>gorithm produces the quantization levels and<br />
threshold of the Lloyd quantization [33].<br />
i) Hopfield Neur<strong>al</strong> N<strong>et</strong>work<br />
A tot<strong>al</strong>ly distinct approach to the quantization is proposed<br />
by Qiu. No direct quantization threshold is given, but instead a<br />
Hopfield neur<strong>al</strong> n<strong>et</strong>work outputs for each input pixel, wh<strong>et</strong>her<br />
it should be quantized to a or b. This m<strong>et</strong>hod gives near to 7 %<br />
b<strong>et</strong>ter MSE-v<strong>al</strong>ues compared to the m<strong>et</strong>hod AMBTC. It has<br />
<strong>al</strong>so been shown that the result is not optim<strong>al</strong>, but gives<br />
virtu<strong>al</strong>ly similar outcome to the Lloyd quantization [34].<br />
Simulation results and an<strong>al</strong>ysis [64] of different<br />
quantization m<strong>et</strong>hods are shown in Table I and Table II.<br />
Initi<strong>al</strong>ly we encode the test images which mentioned in Table I<br />
and Table II by above discussed m<strong>et</strong>hods. Quantization levels<br />
(a,b) are encoded by 8+8 bits, and the bit plane is stored as<br />
such. The qu<strong>al</strong>ity of the reconstructed images is shown in<br />
Table I and Table II. The results indicate that the difference<br />
b<strong>et</strong>ween the origin<strong>al</strong> second moment preserving and the MSEoptim<strong>al</strong><br />
quantization is remarkable. Magnifications of the<br />
reconstructed images are shown in Figure 2. The coding<br />
artifacts are clearly seen in the magnifications; however, it is<br />
<strong>al</strong>most impossible to observe any subjective qu<strong>al</strong>ity<br />
differences b<strong>et</strong>ween the quantization m<strong>et</strong>hods.<br />
The MSE-v<strong>al</strong>ues are compared with the corresponding<br />
MSE-optim<strong>al</strong> quantization v<strong>al</strong>ues in Figure 2. The m<strong>et</strong>hods<br />
that safeguard some s<strong>et</strong> of moments are worst in the MSEsense.<br />
The best candidates aside MSE-optim<strong>al</strong> quantization are<br />
the Lloyd and the one which uses GB-heuristics as a threshold.<br />
The latter is <strong>al</strong>so easy to c<strong>al</strong>culate, and is thus a good<br />
<strong>al</strong>ternative to the MSE-optim<strong>al</strong> quantization. In the following<br />
sections, the MSE-optim<strong>al</strong> quantization will be the point of<br />
comparisons, unless otherwise noted.<br />
The basic idea behind the origin<strong>al</strong> BTC was to preserve<br />
certain moments. Despite the fact that the moment preserving<br />
quantizers can be written in closed form, the resulting<br />
quantization levels are only integer approximations of the<br />
formulas, and thus distorted by rounding errors. Figure 4 and 5<br />
demonstrate how well the two first moments are actu<strong>al</strong>ly<br />
preserved in different quantization systems (non black sections<br />
in the columns).<br />
The illustrations show that the moments are not preserved<br />
precisely in any m<strong>et</strong>hod. In fact, the moment preserving<br />
quantization does not preserve the first moment (mean) any<br />
b<strong>et</strong>ter than the other quantization m<strong>et</strong>hods (except the MAEoptim<strong>al</strong><br />
quantization). It will, however, preserve the second<br />
moment somewhat b<strong>et</strong>ter, but the difference is hardly<br />
significant.<br />
Moreover, if the quantization data and/or bit plane is<br />
further compressed, the moments are even less accurately<br />
preserved. In Figure 4 and 5, there are <strong>al</strong>so errors originating<br />
from quantization of (a,b) to 6+6 bits (gray sections in the<br />
columns) and from 50 % interpolation of the bit plane (dark<br />
sections in the columns). Despite <strong>al</strong>l of that, the greatest<br />
deficiency of the moment preserving quantization is that it<br />
does not even try to preserve the moments beyond the block<br />
boundaries.<br />
Figure 2(a).Origin<strong>al</strong><br />
bpp=8.00 MSE=0<br />
Figure 2(c).3rd moment<br />
bpp=2.00 MSE=42.11<br />
Figure 2(e).GB+mean<br />
bpp=2.00 MSE=37.05<br />
Figure 2(g).MSE-optim<strong>al</strong><br />
bpp=2.00 MSE=35.54<br />
Figure 2(b).2nd moment<br />
bpp=2.00 MSE=43.76<br />
Figure 2(d).Absolute moment<br />
bpp=2.00 MSE=40.51<br />
Figure 2(f).Lloyd<br />
bpp=2.00 MSE=36.83<br />
Figure 2(h).MAE-optim<strong>al</strong><br />
bpp=2.00 MSE=40.21<br />
Figure2. MSE and bpp v<strong>al</strong>ues for different moment preserving <strong>al</strong>gorithms<br />
IJCTA | JULY-AUGUST 2011<br />
Available online@www.ijcta.com<br />
1102
<strong>Aditya</strong> <strong>kumar</strong> <strong>et</strong> <strong>al</strong>, <strong>Int</strong>. J. <strong>Comp</strong>. <strong>Tech</strong>. <strong>Appl</strong>., <strong>Vol</strong> 2 (4), 1099-1117<br />
ISSN:2229-6093<br />
compact representation for the block without any redundancy<br />
and cross correlations.<br />
In the origin<strong>al</strong> moment preserving BTC) the quantization<br />
data are represented by the pair (x̅, σ). This is not the case for<br />
the other quantization m<strong>et</strong>hods. For example AMBTC uses the<br />
mean and the absolute first centr<strong>al</strong> moment [30] [31]. If the<br />
heuristic threshold selection of [25] is applied, then one has to<br />
code the threshold and deviation corresponding to it:-<br />
σ = ∑<br />
<br />
<br />
x −<br />
<br />
x <br />
(24)<br />
Figure3. MSE-deficiency compared to the MSE-optim<strong>al</strong> quantization [64].<br />
Figure4. Error in the 1st moments of the reconstructed Lena [64]<br />
Figure5. Error in the 2nd moments of the reconstructed Lena [64]<br />
B. Coding the Quantization Data<br />
The second task in the structured BTC-<strong>al</strong>gorithm is to<br />
encode the given quantization data of a block. Before this can<br />
be done, an important choice must be made, which data should<br />
be coded. There are two distinct approaches for sending the<br />
quantization data to the decoder.<br />
In the first approach the quantization data is expressed by<br />
two statistic<strong>al</strong> v<strong>al</strong>ues, representing the mid-v<strong>al</strong>ue and variation<br />
quantity of a block (usu<strong>al</strong>ly mean and standard deviation). It is<br />
known that the dispersion of the variation quantity is sm<strong>al</strong>ler<br />
than the dispersion of individu<strong>al</strong> pixel intensities, and sm<strong>al</strong>l<br />
v<strong>al</strong>ues occur more frequently than large ones. This gives a<br />
A demerit to this approach is that the quantization levels<br />
are c<strong>al</strong>culated at the decoding phase from the quantized v<strong>al</strong>ues<br />
of (x̅, σ) containing rounding errors. Thus extra degradation is<br />
caused by the coding phase.<br />
The other approach is to c<strong>al</strong>culate the quantization levels<br />
(a, b) <strong>al</strong>ready at the encoding phase and transmit them. In this<br />
way one can minimize both the quantization error and the<br />
computation needed at the decoding phase. They are <strong>al</strong>so<br />
independent of the quantization m<strong>et</strong>hod selected. The pair (a,<br />
b) contains redundancy, which can be easily removed by a<br />
suitable prediction and coding technique.<br />
Sever<strong>al</strong> m<strong>et</strong>hods for coding the quantization data will be<br />
presented next.<br />
a) Fixed Number of Bits<br />
The most straightforward approach is to code the<br />
quantization data by 8+8 bits as such. No extra computations<br />
are then needed and direct access to the coded file is r<strong>et</strong>ained.<br />
It has been observed [20], that the decrease in the qu<strong>al</strong>ity of<br />
the image will be sm<strong>al</strong>l if (x̅, σ) is quantized to 6+4 bits. Here<br />
one can simply <strong>al</strong>locate the v<strong>al</strong>ues evenly within the ranges x̅∈<br />
[0,255], and σ∈ [0,127] (or even truncate the standard<br />
deviation to the range σ∈ [0, 63]). In the same way, less than<br />
8+8 bits can be <strong>al</strong>located for the pair (a, b). Our experiments<br />
indicate that the MSE-v<strong>al</strong>ue of the test image Lena increases<br />
only circa 3 % when using 6+6 bits in the quantization.<br />
b) Joint Quantization<br />
He<strong>al</strong>y and Mitchell [32] [33] improved the 6+4 bits coding<br />
system by applying joint quantization to (x, σ) with 10 bits.<br />
Here the accuracy of x depends inversely on the v<strong>al</strong>ue of σ.<br />
For example if σ=4 there are 10-3=7 bits left to code x. The<br />
technique still gives the same bit rate, but it has been argued<br />
that the quantization error is more visible to the human eye in<br />
the low variance regions, and a more accurate representation is<br />
needed in these.<br />
The concept joint quantization was origin<strong>al</strong>ly introduced<br />
by Mitchell and Delp [35]. They used the m<strong>et</strong>hod for 16- and<br />
32-level images by using 6 and 7 bits for the pair (x, σ),<br />
respectively. See Figure 6.<br />
c) Vector Quantization<br />
The use of vector quantization [28] [29] for coding the bit<br />
plane was first proposed by Udpikar and Raina [61]. They<br />
used a variant of AMBTC [32] that stores a compressed block<br />
in the form (x̅ , x , B). The decoder replaces each 0-bit of B by<br />
x and each 1-bit by x,<br />
<br />
IJCTA | JULY-AUGUST 2011<br />
Available online@www.ijcta.com<br />
1103
<strong>Aditya</strong> <strong>kumar</strong> <strong>et</strong> <strong>al</strong>, <strong>Int</strong>. J. <strong>Comp</strong>. <strong>Tech</strong>. <strong>Appl</strong>., <strong>Vol</strong> 2 (4), 1099-1117<br />
ISSN:2229-6093<br />
where<br />
x = x + . x̅ − x <br />
<br />
<br />
(25)<br />
By applying vector quantization (VQ) to the pair (x, x ),<br />
the bit rate can be reduces to 8 bits for (x, x) with the cost of<br />
increased MSE-v<strong>al</strong>ue. The m<strong>et</strong>hod, however, does not<br />
consider interblock correlations, i.e. the dependencies b<strong>et</strong>ween<br />
neighboring blocks.<br />
Weitzman and H.B. Mitchell [46] apply predictive vector<br />
quantization for reducing the inter block redundancy. Instead<br />
of coding (a, b), they apply VQ to the prediction errors (ea i ,<br />
eb i ):-<br />
ea = a − 2a + 2a + a /5 (26)<br />
eb = b − 2b + 2b + b /5 (27)<br />
Here a , a , and a refer to the a-v<strong>al</strong>ues of<br />
the neighboring blocks that have <strong>al</strong>ready been coded.<br />
Figure6. Joint Quantization of x̅ and σ [64]<br />
Another approach was proposed by Alcaim and Oliveira<br />
[47]. They encode x and σ separately, making two subsample<br />
images, one from the x-v<strong>al</strong>ues and another from the σ-v<strong>al</strong>ues<br />
of the blocks. These subsample images are decomposed into<br />
2*2 blocks which are then coded by VQ. In this way one can<br />
reduce both the interblock and intrablock redundancy.<br />
d) DISCRETE COSINE TRANSFORM<br />
Wu and Coll [36] propose the compression of the pair (a,<br />
b) by the use of discr<strong>et</strong>e cosine transform (DCT). Two<br />
subsample images are formed, one from the a-v<strong>al</strong>ues and<br />
another from the b-v<strong>al</strong>ues of the blocks. These subsample<br />
images contain one pixel per block of the origin<strong>al</strong> image. They<br />
are then coded by the adaptive DCT presented in [20] [31].<br />
D<strong>et</strong>ails of the subsample images must be preserved as<br />
accurately as possible, since they have an influence on a large<br />
number of pixels in the reconstructed image.<br />
Here the second subsample image is not formed from the<br />
b-v<strong>al</strong>ues, but from the (b-a)-differences instead. If one wants<br />
to r<strong>et</strong>ain the same MSE-level, the differences must be<br />
c<strong>al</strong>culated at the coding phase by b-a', where a' is the<br />
reconstructed a-v<strong>al</strong>ue after inverse DCT.<br />
The use of a computation<strong>al</strong>ly demanding <strong>al</strong>gorithm like<br />
DCT is justified by the fact that the subsample images are only<br />
1/16 of size of the origin<strong>al</strong> image, for the block size 4*4.<br />
When the compression ratio of the DCT is predefined to 3:1,<br />
the tot<strong>al</strong> bit rate of a and b is 0.33 bpp at the cost of higher<br />
MSE.<br />
e) Lossless Coding<br />
The two subsample images can be compressed by sever<strong>al</strong><br />
different image coding <strong>al</strong>gorithms. As stated earlier, it is<br />
important that the information of the subsample images is<br />
preserved as far as possible. Therefore the use of a lossless<br />
image compression <strong>al</strong>gorithm is apparently a good choice.<br />
Two variants of this type, and another related m<strong>et</strong>hod.<br />
We consider next a new approach for coding the two<br />
subsample images. We apply FELICS coding (Fast and<br />
Efficient Lossless Image <strong>Comp</strong>ression System) by Howard<br />
and Vitter [25]. The m<strong>et</strong>hod is highly applicable to BTC<br />
because of its practic<strong>al</strong> usefulness, and it is <strong>al</strong>so one of the<br />
most efficient lossless m<strong>et</strong>hods known.<br />
FELICS coding uses the information of two adjacent pixels<br />
when coding the current one. These are the one to the left of<br />
the current pixel, and the one above it. Denote the v<strong>al</strong>ues of<br />
the neighboring pixels by L and H so that L is the one which is<br />
sm<strong>al</strong>ler. Howard and Vitter [25] have found that the<br />
probability of individu<strong>al</strong> pixels obeys the distribution given in<br />
Figure 7. Our experiments show that the same distribution<br />
holds very well for the pixels of the subsample images, too.<br />
The coding scheme is as follows: A code bit indicates<br />
wh<strong>et</strong>her the actu<strong>al</strong> v<strong>al</strong>ue f<strong>al</strong>ls into the in-range. If so, an<br />
adjusted binary coding is applied. Here the hypothesis is that<br />
the in-range v<strong>al</strong>ues are uniformly distributed. Otherwise the<br />
above/below-range decision requires another code bit, and the<br />
v<strong>al</strong>ue is then coded by Rice coding [64] with adaptive k-<br />
param<strong>et</strong>er selection. For d<strong>et</strong>ails of FELICS coding, see [64].<br />
We can improve the coding of b if we remember that a is<br />
<strong>al</strong>ready known at the moment of coding b, and that b≥a. Thus,<br />
if a f<strong>al</strong>ls into the same interv<strong>al</strong> as b, the actu<strong>al</strong> range of b can<br />
be reduced so that only the potenti<strong>al</strong> v<strong>al</strong>ues are considered.<br />
For example if a belongs to the in-range of b, the interv<strong>al</strong> is<br />
reduced by changing the lower bound L to a. The<br />
improvement, however, remains margin<strong>al</strong>.<br />
The results of FELICS are quite similar compared to the<br />
results of the predictive entropy coding scheme, but it is fast<br />
and much less involved.<br />
Figure7. Probability distribution of intensity v<strong>al</strong>ues<br />
C. Bit Plane Reduction<br />
In the basic BTC m<strong>et</strong>hod, both quantization data and the<br />
bit plane of a block require 16 bits each. Sever<strong>al</strong> m<strong>et</strong>hods for<br />
IJCTA | JULY-AUGUST 2011<br />
Available online@www.ijcta.com<br />
1104
<strong>Aditya</strong> <strong>kumar</strong> <strong>et</strong> <strong>al</strong>, <strong>Int</strong>. J. <strong>Comp</strong>. <strong>Tech</strong>. <strong>Appl</strong>., <strong>Vol</strong> 2 (4), 1099-1117<br />
ISSN:2229-6093<br />
reducing the bits needed for the quantization data was<br />
presented in vector quantization. A larger reduction is still<br />
possible at the cost of decreased image qu<strong>al</strong>ity. In this section,<br />
we will consider sever<strong>al</strong> attempts to reduce the size of the bit<br />
plane.<br />
a) Skipping the Bit Plane<br />
If the contrast of x i -v<strong>al</strong>ues is very sm<strong>al</strong>l one can except the<br />
storing of B and thus implicitly code <strong>al</strong>l pixels by x̅ . The<br />
criterion for using this one-level quantization varies. Mitchell<br />
and Delp [35] omit the bit plane if σ
ISSN:2229-6093<br />
<strong>Aditya</strong> <strong>kumar</strong> <strong>et</strong> <strong>al</strong>, <strong>Int</strong>. J. <strong>Comp</strong>. <strong>Tech</strong>. <strong>Appl</strong>., <strong>Vol</strong> 2 (4), 1099-1117<br />
opening and roots of closing are adaptively used in the<br />
m<strong>et</strong>hod. See Figure 10.<br />
Figure10. Roots of the Filtering [64]<br />
d) Vector Quantization<br />
Sever<strong>al</strong> approaches have been proposed for reducing the<br />
volume of the bit plane by the use of vector quantization or a<br />
similar ad hoc technique [48]. The basic idea of VQ is to<br />
select a sm<strong>al</strong>l s<strong>et</strong> of representative vectors (binary blocks) and<br />
to code <strong>al</strong>l possible vectors by an index. The codebook is<br />
generated off-line on the basis of training vectors by using a<br />
suitable <strong>al</strong>gorithm, like the gener<strong>al</strong>ized Lloyd <strong>al</strong>gorithm [28].<br />
Then the same codebook is used for whatever images are to be<br />
coded. Adaptive vector quantization [42] [43] was considered<br />
by Wu and Coll [34], but they prefer the non-adaptive,<br />
univers<strong>al</strong> codebook VQ, because of the speed.<br />
The representative for a bit plane can be chosen by a tot<strong>al</strong><br />
search, i.e. by testing each candidate code vector and selecting<br />
the one which minimizes the given distortion function. In this<br />
way, the result of the search is <strong>al</strong>ways optim<strong>al</strong> (corresponding<br />
to the distortion function) but it is impractic<strong>al</strong> for large<br />
codebooks. The decoding, however, is <strong>al</strong>ways quick.<br />
A loc<strong>al</strong>ized search was proposed by Weitzman and H.B.<br />
Mitchell [40]. They divide the codebook into sever<strong>al</strong><br />
overlapping sub-codebooks. For each bit plane, its<br />
combination index is c<strong>al</strong>culated. It d<strong>et</strong>ermines the subcodebook<br />
in which the search is performed. Even if the search<br />
m<strong>et</strong>hod is not extensive anymore, the correct code vector<br />
(corresponding to the extensive search) will still be found with<br />
a very high probability (> 99%) according to Michell [20].<br />
The performance of a search m<strong>et</strong>hod depends not only on<br />
the organization of the codebook, but <strong>al</strong>so on the selected<br />
distortion function. MSE is the most obvious measure,<br />
however, it cannot be c<strong>al</strong>culated independently from the<br />
quantization data (a, b) [17]. Most of the existing m<strong>et</strong>hods<br />
thus use Hamming distance, i.e. the number of positions in<br />
which the elements of the two binary vectors differ, as a<br />
distortion measure [48]. If there are two or more code vectors<br />
with the same Hamming distance from the origin<strong>al</strong> bit vector,<br />
the one with the same number of 1-bits is chosen. The<br />
advantage of Hamming distance is that it can be implemented<br />
by look-up table (LUT).<br />
L<strong>et</strong>’s suppose the size of the codebook was 256 and 128,<br />
so the bit rate of the bit plane was respectively 0.5 and 0.44.<br />
Weitzman and H.B. Mitchell [42] [43] proposed the use of<br />
sever<strong>al</strong> codebooks of different sizes. The size of a codebook<br />
depends on the contrast of a block (measured by b-a): the<br />
greater the contrast the larger the codebook. Nasiopoulos [48]<br />
coded only blocks whose contrast is b<strong>et</strong>ween a given range<br />
(not too large but not too low either) by VQ. The codebook<br />
was formed so that edge blocks are well represented. This <strong>al</strong>so<br />
reduces the so-c<strong>al</strong>led staircase effect, which is otherwise<br />
impaired by VQ.<br />
A classified VQ <strong>al</strong>gorithm is proposed by Efrati <strong>et</strong> <strong>al</strong>. [52]<br />
for avoiding the staircase effect. BTC <strong>al</strong>gorithm with VQ is<br />
used for the low contrast blocks, and a three-level quantizer is<br />
applied for the high contrast blocks. Thus, the <strong>al</strong>gorithm uses<br />
two different codebooks, one for the low contrast and another<br />
for the high contrast blocks.<br />
e) <strong>Int</strong>erpolation<br />
All the m<strong>et</strong>hods discussed above were based on the idea<br />
that the bit plane is only parti<strong>al</strong>ly coded, thus omitting a part<br />
of it, or by selecting a representative for the bit plane (or for its<br />
rows) from a sm<strong>al</strong>ler subs<strong>et</strong>, which can either be c<strong>al</strong>led a root<br />
sign<strong>al</strong> s<strong>et</strong> or a codebook. Therefore the reconstructed image<br />
contains pixels where an a-v<strong>al</strong>ue has been changed to a b-<br />
v<strong>al</strong>ue or vice versa. The cost of these errors may be high in the<br />
MSE-sense, if the contrast in the block is large. To avoid this<br />
problem, a tot<strong>al</strong>ly different approach is taken in the following<br />
m<strong>et</strong>hod.<br />
Zeng's interpolative BTC [41] [42] is similar to the<br />
prediction technique (m<strong>et</strong>hod B) in the sense that h<strong>al</strong>f of the<br />
bit plane is omitted. At the decoding phase the parti<strong>al</strong> image is<br />
first reconstructed and then the missing pixel v<strong>al</strong>ues are<br />
interpolated on the basis of the existing ones. Zeng has<br />
proposed two variants of the m<strong>et</strong>hod: one which codes 50 %<br />
(IBTC-1) and another which codes 25 % of the bits (IBTC-2).<br />
Figure11. <strong>Int</strong>erpolation patterns of IBTC-1 and IBTC-2 [64]<br />
The interpolation is done by stack filters designed to be<br />
optim<strong>al</strong> in the MAE-sense [38] [41] or by using median filters<br />
[42]. The latter is recommended because of the ease of<br />
implementation. In the IBTC-1 four adjacent coded pixels are<br />
used to c<strong>al</strong>culate the intensity of the current one.<br />
y , = median y , , y , , y , , y , , . y , + y , +<br />
y , + y , (29)<br />
IJCTA | JULY-AUGUST 2011<br />
Available online@www.ijcta.com<br />
1106
<strong>Aditya</strong> <strong>kumar</strong> <strong>et</strong> <strong>al</strong>, <strong>Int</strong>. J. <strong>Comp</strong>. <strong>Tech</strong>. <strong>Appl</strong>., <strong>Vol</strong> 2 (4), 1099-1117<br />
ISSN:2229-6093<br />
In the IBTC-2 a two-phase interpolation is used. At the<br />
first phase the following median filter is used, where each<br />
pixel is interpolated on the basis of the coded pixels.<br />
y , = median y , , y , , y , , y , , 1 4 . y ,<br />
+ y , + y , + y , (30)<br />
Although the hierarchic<strong>al</strong> decomposition seems to be an<br />
excellent idea, it has some drawbacks <strong>al</strong>so. Sever<strong>al</strong> nonhierarchic<strong>al</strong><br />
m<strong>et</strong>hods are preserved direct access to the<br />
compressed file, but a straightforward implementation of<br />
HBTC does not. This <strong>al</strong>so means that the fault tolerance has<br />
decreased. Moreover, large uniform areas may cause blocky<br />
appearance in the reconstructed image.<br />
At the second phase, the median filter of (2.30) is used.<br />
Note that the interpolation is a separate phase of the decoding<br />
process and therefore the block boundaries do not disturb the<br />
reconstruction of the missing pixels.<br />
f) Entropy Coding<br />
Since the bit plane can be considered as a binary image,<br />
one might expect that a binary image coding m<strong>et</strong>hod could be<br />
used. On the other hand, the BTC-process ensures that the<br />
numbers of 0 and 1 bits are <strong>al</strong>most equ<strong>al</strong> and the bits are<br />
relatively evenly distributed <strong>al</strong>l over the bit plane. Therefore<br />
simple coding m<strong>et</strong>hods, like run length coding [44], are unlike<br />
to give any compression.<br />
In the entropy based bit plane compression, the image (the<br />
bit plane) is preceded in row major order from left to right. A<br />
high degree Markov model is applied for each pixel. The v<strong>al</strong>ue<br />
of a pixel is predicted by the 7-bit context template and then<br />
coded by arithm<strong>et</strong>ic coding according to its probability. QMcoder<br />
[45, 46] was used as the arithm<strong>et</strong>ic coding component.<br />
The size and shape of the template have major effect on the<br />
coding efficiency when compressing binary images and b<strong>et</strong>ter<br />
results can be achieved with larger templates. However, this<br />
does not necessarily apply to the block truncation coding. The<br />
problem here is the block wise quantization according to the<br />
mean x of each individu<strong>al</strong> block. This destroys the correlation<br />
of neighboring bits belonging to two different blocks.<br />
D. Hierarchic<strong>al</strong> Decomposition of the Image<br />
A natur<strong>al</strong> extension of the single-level BTC is to use a<br />
hierarchy of blocks [44] [53]. In hierarchic<strong>al</strong> block truncation<br />
coding (HBTC) the processing begins with a large block size<br />
(m 1 *m 1 ). If variance σ2 of a block is less than a predefined<br />
threshold vth the block is coded by a BTC-variant. Otherwise<br />
it is divided into four subblocks and the same process is<br />
repeated until the variance threshold criterion is m<strong>et</strong>, or the<br />
minim<strong>al</strong> block size (m 2 *m 2 ) is reached.<br />
By adjusting the variance threshold v th , and the block sizes<br />
m 1 , m 2 , one can direct the benefit of the hierarchic<strong>al</strong><br />
decomposition to achieve a low bit rate or high image qu<strong>al</strong>ity.<br />
The high MSE-v<strong>al</strong>ues rarely originate from the largest<br />
blocks unless the variance threshold is too high. Therefore, if a<br />
low MSE-v<strong>al</strong>ue is the primary criterion of the image qu<strong>al</strong>ity,<br />
the maxim<strong>al</strong> block size of 32*32 is recommended. The<br />
selection for the sm<strong>al</strong>lest block size m 2 causes a more dramatic<br />
change in MSE. If one wants to achieve very high qu<strong>al</strong>ity<br />
images the sm<strong>al</strong>lest block size must be s<strong>et</strong> to m 2 =2. The<br />
variance threshold v th can be considered as a fine tuning<br />
param<strong>et</strong>er.<br />
Figure12. Bit plane of Lena by BTC [64]<br />
Figure13. Bit plane of Lena by HBTC [64]<br />
Figure14. Bit plane of Lena by HBTC<br />
IJCTA | JULY-AUGUST 2011<br />
Available online@www.ijcta.com<br />
1107
<strong>Aditya</strong> <strong>kumar</strong> <strong>et</strong> <strong>al</strong>, <strong>Int</strong>. J. <strong>Comp</strong>. <strong>Tech</strong>. <strong>Appl</strong>., <strong>Vol</strong> 2 (4), 1099-1117<br />
ISSN:2229-6093<br />
E. Extensions of BTC<br />
This section are included, sever<strong>al</strong> extern<strong>al</strong> pre- and postprocessing<br />
techniques will be discussed. Some of them aim at<br />
an improved visu<strong>al</strong> qu<strong>al</strong>ity of the reconstructed image, while<br />
others simply strive for lower bit rates. BTC application to<br />
color images and video compression will be shortly discussed.<br />
a) Smoothing Filter<br />
A simple smoothing operation of the origin<strong>al</strong> image is<br />
proposed by Mitchell and Delp [35]. The purpose of the<br />
smoothing is to reduce noise in the origin<strong>al</strong> image which<br />
causes pixel v<strong>al</strong>ues near the threshold to be quantized to the<br />
wrong v<strong>al</strong>ue. Each pixel x i,j of a block is replaced by a<br />
weighted sum of itself and the four adjacent pixels.<br />
<br />
x , = 0.6 . x , + 0.1. x , + x , + x , + x , (31)<br />
b) Blocky Appearance<br />
Blocky appearance is a common feature of block coding<br />
techniques like BTC. The relatively sm<strong>al</strong>l size of blocks (4*4)<br />
keeps the effect <strong>al</strong>most unnoticeable for human eyes.<br />
However, when large block sizes are applied in the<br />
hierarchic<strong>al</strong> variant, the effect becomes more visible.<br />
this difference image. Both BTC and transform codes are sent<br />
to the decoder with bit rate 1.63 + 0.25 = 1.88 bpp.<br />
The hybrid m<strong>et</strong>hod was argued for using the different<br />
properties of the m<strong>et</strong>hods. While BTC preserves edges, they<br />
tend to be ragged. On the other hand, transform coding usu<strong>al</strong>ly<br />
produces smooth edges. In the hybrid formulation, the aim is<br />
to improve visu<strong>al</strong> qu<strong>al</strong>ity of the image. At the same time,<br />
MSE-v<strong>al</strong>ue is reported to decrease [20].<br />
Figure16. Image qu<strong>al</strong>ity versus block size [64]<br />
d) Differenti<strong>al</strong> Pulse Code Modulation<br />
Differenti<strong>al</strong> pulse code modulation (DPCM) is a wellknown<br />
sign<strong>al</strong> coding technique and it has been widely used in<br />
image compression [54] [55]. In DPCM, each pixel v<strong>al</strong>ue is<br />
first predicted on the basis of previously coded pixels. The<br />
result of the prediction is then subtracted from the actu<strong>al</strong> pixel<br />
v<strong>al</strong>ues and the difference e ij =x ij -y ij is coded.<br />
Figure15. Bit <strong>Comp</strong>ression efficiency versus block size [64]<br />
For reducing the blocky appearance Nasiopoulos [48]<br />
proposed the use of smoothing filter as a post processing<br />
technique. The reconstructed image is scanned using an 8*8-<br />
window. If the contrast in the window is sm<strong>al</strong>ler than a<br />
predefined smoothness criterion, the area is smoothed by a 3*3<br />
averaging filter. Here each pixel y i,j of the window is replaced<br />
by the v<strong>al</strong>ue y' i,j .<br />
<br />
= ∑ ∗ ∑ y ,<br />
y ,<br />
<br />
<br />
(32)<br />
The m<strong>et</strong>hod takes care of the blocky appearance of the<br />
reconstructed image, but the smoothed regions now look flat<br />
and artifici<strong>al</strong>. In order to give back to the image the lost<br />
texture, Gaussian random noise is added to the averaged areas.<br />
c) Discr<strong>et</strong>e Cosine Transform<br />
A hybrid of BTC and DCT is proposed by Delp and<br />
Mitchell [20]. First a highly compressed image is obtained by<br />
taking the two-dimension<strong>al</strong> discr<strong>et</strong>e cosine transform over<br />
16*16 pixel blocks. Only the eight non-dc coefficients in the<br />
low frequency section of each block are r<strong>et</strong>ained. A difference<br />
image is constructed by subtracting the transform coded image<br />
from the origin<strong>al</strong>. The basic BTC <strong>al</strong>gorithm is then applied to<br />
Figure17. Differenti<strong>al</strong> Pulse Code Modulation<br />
Delp and Mitchell [56] propose a composite m<strong>et</strong>hod of<br />
DPCM and BTC. First an open-loop prediction is performed to<br />
construct the mean v<strong>al</strong>ue and standard deviation of the<br />
differences (e, σ e ). In this stage, quantization levels a and b are<br />
<strong>al</strong>so c<strong>al</strong>culated, because they are needed for the predictor.<br />
Then the quantization procedure is performed by applying<br />
DPCM with a closed-loop prediction. Prediction function is as<br />
follows.<br />
x̅ = 0.8. y , − 0.6. y , + 0.8. y , (33)<br />
Because the ranges of e and σ e are sm<strong>al</strong>ler than those of x<br />
and σ, they can be coded with fewer bits. With the block size<br />
8*8, Delp and Mitchell have found that 4+3 bits suffice;<br />
however, the d<strong>et</strong>ailed bit <strong>al</strong>location table is not given.<br />
IJCTA | JULY-AUGUST 2011<br />
Available online@www.ijcta.com<br />
1108
<strong>Aditya</strong> <strong>kumar</strong> <strong>et</strong> <strong>al</strong>, <strong>Int</strong>. J. <strong>Comp</strong>. <strong>Tech</strong>. <strong>Appl</strong>., <strong>Vol</strong> 2 (4), 1099-1117<br />
ISSN:2229-6093<br />
e) N-Level Quantization<br />
The basic idea of block truncation coding, two-level<br />
quantization can be derived to a gener<strong>al</strong> n-level quantizer.<br />
Delp and Mitchell have proposed a moment preserving<br />
quantizer, which is shown to be related to the Gauss-Jacobi<br />
mechanic<strong>al</strong> quadrature problem [33] [43]. MSE-optim<strong>al</strong><br />
quantizer has been given by Max [35]. These quantizer,<br />
however, presuppose that the probability distribution of the<br />
sign<strong>al</strong> is either uniform, Gaussian, or Laplacian. In Lloyd's<br />
quantizer [33] <strong>al</strong>l quantization levels are given by an iterative<br />
<strong>al</strong>gorithm.<br />
Mon<strong>et</strong> and Dubois [58] have proposed a uniform<br />
quantization m<strong>et</strong>hod in which the overhead information of a<br />
block is first coded, consisting of the pair (x min , x max ). The<br />
pixel v<strong>al</strong>ues of a block are then sc<strong>al</strong>ed and norm<strong>al</strong>ized within<br />
this interv<strong>al</strong>. They are fin<strong>al</strong>ly quantized so that the number of<br />
quantization levels in a block is fixed regardless of the<br />
variance of the block.<br />
Quantizer<br />
TABLE III Properties of Different Quantizers [64]<br />
Quantization<br />
param<strong>et</strong>ers:<br />
Quant.<br />
levels:<br />
Overhead<br />
Quant. Quant.<br />
param.: matrices<br />
BTC (a,b) 2 2 1<br />
3-level (a,b) 3 2 log 3<br />
BTC<br />
4-level (a,b) 4 2 2<br />
BTC<br />
n-level One v<strong>al</strong>ue N N log N<br />
BTC<br />
Block AQ (x max, x min) N 2 log N<br />
Lyod Quan. level N N log N<br />
More specific approaches have been taken by sever<strong>al</strong><br />
authors concentrating on the 3-level quantization [25] [52].<br />
Goeddel and Bass [19] proposed a 3-level quantizer to be used<br />
<strong>al</strong>ong with the origin<strong>al</strong> BTC. For each block, both<br />
quantizations are d<strong>et</strong>ermined, but only the one with the sm<strong>al</strong>ler<br />
MSE-v<strong>al</strong>ue is sent to decoder. Ronsin and DeWitte [22] [28]<br />
use both 3- and 4-level quantization. They later developed<br />
another <strong>al</strong>gorithm, in which the edge blocks are coded by a 3-<br />
level and the non-edge blocks by a 2-level quantizer [59].<br />
Vector quantization of the quantization matrix is<br />
performed in <strong>al</strong>l of these approaches. The codebook is<br />
designed to contain edge blocks [24]; or the end points of the<br />
edges are coded by an edge following <strong>al</strong>gorithm [59]. Wang<br />
and Mitra proposed a more sophisticated <strong>al</strong>gorithm by using<br />
block pattern models [60].<br />
Ceng and Tsai [61] proposed the use of a 4-level quantizer<br />
without increasing the number of param<strong>et</strong>ers. All four<br />
quantization levels can still be derived from the pair (a,b). Wu<br />
and Coll [62] have given a non-uniform n-level quantizer,<br />
where each of the quantization levels are iteratively<br />
d<strong>et</strong>ermined aiming at minim<strong>al</strong> MAE. They <strong>al</strong>so introduced a<br />
predictive entropy coding scheme for coding the overhead in<br />
the speci<strong>al</strong> case of a 4-level quantizer. Uhl [62] has extended<br />
the idea of multilevel quantization and proposed a hybrid<br />
<strong>al</strong>gorithm. The blocks are classified and for each class, a<br />
different quantization is applied. They use 1, 2, 4, and 8-level<br />
quantizers, and a copy operation depending on the type of<br />
block.<br />
f) Color Images<br />
The most common representation of color images, RGB,<br />
divides the image into three separate color planes, namely red,<br />
green, and blue. This color system is based on the trichromatic<br />
nature of the human vision [6]. There exists a high degree of<br />
correlation b<strong>et</strong>ween the planes R, G and B, and therefore a<br />
color space conversion from RGB to YUV (or YIQ) is usu<strong>al</strong>ly<br />
performed [21]. Y represents the luminance of the image,<br />
while U, V (I,Q) consists of the color information, i.e.<br />
chrominance.<br />
Color images are considered as a three-component<br />
gener<strong>al</strong>ization of gray sc<strong>al</strong>e images, and thus they can be<br />
compressed by extending the existing <strong>al</strong>gorithms. Lema and<br />
Mitchell [31] propose an <strong>al</strong>gorithm in which the three<br />
components of the YIQ-space are separately coded by the<br />
absolute moment BTC. As the human visu<strong>al</strong> system is most<br />
sensitive to the variations of Y, and less to I and least to Q, the<br />
chrominance planes are sub sampled. It means, that only the<br />
average v<strong>al</strong>ue of each 2*2-block in the I-plane and the average<br />
v<strong>al</strong>ue of each 4*4-block in the Q-plane are r<strong>et</strong>ained. In the<br />
decompression phase, the sizes of the planes are recovered by<br />
using bilinear interpolations. The m<strong>et</strong>hod yields a bit rate of<br />
2.13 (1.62+0.41+0.10) bpp.<br />
Wu and Coll [67] proposed the use of a single bit map for<br />
<strong>al</strong>l three components, given by the quantization of the Y-<br />
component. The quantization levels of each component are<br />
further compressed by DCT. High-frequency components of<br />
the DCT domain are discarded in the chrominance<br />
components so, that it correspond to a 3*3 averaging filter for<br />
the I-sign<strong>al</strong> and a 5*5 averaging filtering for the Q-sign<strong>al</strong>.<br />
Sever<strong>al</strong> variations of the basic idea are considered in [68],<br />
including both RGB and YIQ color space images. A bit rate of<br />
1.71 bpp has been reported for the latter.<br />
g) Video Images<br />
A straightforward application of BTC to video images<br />
proposed by He<strong>al</strong>y and Mitchell [69]. They extended the block<br />
to consist of pixels from consecutive frames by using block<br />
size of 4*4*3. A registration <strong>al</strong>gorithm is used for d<strong>et</strong>ecting<br />
glob<strong>al</strong> translations, like camera movements. 2D-BTC is<br />
performed on movement blocks and edge regions. Bit rate can<br />
be reduced by coding only one 2D bit plane for the 3D-blocks.<br />
Further reduction is possible by skipping the coding of one or<br />
two frames, and coding only the blocks consisting of some<br />
activities. All of these ideas are applied adaptively so that<br />
more accurate representation is used whenever necessary for<br />
preserving the image qu<strong>al</strong>ity.<br />
h) Block To Block Adaptive BTC<br />
In an adaptive compression scheme one can change the<br />
compression m<strong>et</strong>hod from block to block. The adaptation may<br />
concern the quantization m<strong>et</strong>hod, bit plane coding m<strong>et</strong>hod, or<br />
both of them. In principle, one could even change the whole<br />
compression system, for example use BTC for one block and<br />
another coding system for the next block. Sever<strong>al</strong> attempts of<br />
IJCTA | JULY-AUGUST 2011<br />
Available online@www.ijcta.com<br />
1109
<strong>Aditya</strong> <strong>kumar</strong> <strong>et</strong> <strong>al</strong>, <strong>Int</strong>. J. <strong>Comp</strong>. <strong>Tech</strong>. <strong>Appl</strong>., <strong>Vol</strong> 2 (4), 1099-1117<br />
ISSN:2229-6093<br />
this kind apply multilevel quantizers <strong>al</strong>ong with the origin<strong>al</strong><br />
BTC.<br />
Table IV<br />
Standard deviation Bit plane coding: Bit rate:<br />
High Full resolution 16 bits<br />
Medium Prediction technique 8 bits<br />
Low Uniform block 0 bits<br />
Griswold [70] vary the quantization system so that a<br />
different s<strong>et</strong> of moments (r'th and 2r'th moments) is preserved.<br />
For each block, the BTC is tentatively applied with <strong>al</strong>l v<strong>al</strong>ues<br />
of r=1, 2, 3, 4, and the one with minimum MSE-v<strong>al</strong>ue is<br />
selected. Quantization levels (a,b) are c<strong>al</strong>culated <strong>al</strong>ready in the<br />
encoding phase, and therefore decoder does not need to know<br />
the choice of r.<br />
Adaptive compression coding (ACC) is proposed by<br />
Nasiopoulos [41]. If the range of a 4*4 block is sm<strong>al</strong>ler than a<br />
given smoothness threshold (2.18, 2.30, or2.42) it is<br />
represented by its average. Blocks that are above the threshold<br />
but below 120 are coded by absolute moment BTC. If the<br />
difference of quantization levels (b-a) is less than or equ<strong>al</strong> to<br />
20, the bit plane is coded by vector quantization including<br />
only edge blocks. Hierarchic<strong>al</strong> decomposition is applied to the<br />
remaining high activity blocks. Sm<strong>al</strong>l blocks are either copied,<br />
or compressed by the one-level quantization. Overhead<br />
information is needed to inform the decoder of the coding<br />
system used.<br />
Figure17. Distribution of 4*4 blocks [64].<br />
Figure18. Average MSE of different blocks<br />
Figure19. Addition<strong>al</strong> MSE contributed by bit plane coding according to the<br />
block types<br />
IV.<br />
COMPARATIVR ANALYSIS ON RECENT BTC ALGORITHM<br />
In this section we discuss recently introduced <strong>al</strong>gorithms<br />
which based on moment preserving criteria and uses in LCD<br />
overdrive like applications.<br />
A. FBTC ALGORITHM<br />
In this section, FBTC (Fast Pixel Grouping <strong>Tech</strong>nique)<br />
above mentioned current scheme is discussed. To clearly<br />
explain the FBTC technique, AMBTC is used here instead of<br />
MPBTC [49] because AMBTC provides b<strong>et</strong>ter image qu<strong>al</strong>ity<br />
and consumes less computation<strong>al</strong> cost. Therefore, pixel<br />
grouping technique [48] can <strong>al</strong>so be employed to improve the<br />
performance of AMBTC.<br />
In this scheme (FBTC), each graysc<strong>al</strong>e image to be<br />
compressed is first partitioned into a s<strong>et</strong> of non-overlapping<br />
image blocks of nxn pixels. Each image block x=<br />
{p 1 ,p 2 ,…..p nxn } is sequenti<strong>al</strong>ly processed. Initi<strong>al</strong>ly, the mean<br />
v<strong>al</strong>ue x of is c<strong>al</strong>culated and stored in x .All pixels with v<strong>al</strong>ues<br />
sm<strong>al</strong>ler than x are classified into the first group G 0 . Otherwise,<br />
they are classified into the second group G 1 . Then, the mean<br />
v<strong>al</strong>ue of pixels in each group is computed and taken as the<br />
quantization level. The quantization levels ql 0 and ql 1 in G 0<br />
and G 1 are c<strong>al</strong>culated according to the following equations,<br />
respectively<br />
ql =<br />
ql = ∑<br />
∑ x <br />
<br />
(34)<br />
x <br />
(35)<br />
In the tradition<strong>al</strong> AMBTC scheme ql 0 and ql 1 are used to<br />
represent pixels in G 0 and G 1 , respectively. Theor<strong>et</strong>ic<strong>al</strong>ly,<br />
pixels in each group should be quite close to the corresponding<br />
quantization level. However, two confusing conditions may be<br />
happened. First, some pixels belonging to G 0 may have their<br />
pixel v<strong>al</strong>ues closer to ql 1 than ql 0 . Second, some pixels in G 1<br />
may have their pixel v<strong>al</strong>ues closer to ql 0 .than ql 1 . When either<br />
one of the two conditions is found, the following processing<br />
steps are iteratively executed in the proposed scheme.<br />
Step 1: Classify <strong>al</strong>l the pixels in x into two groups<br />
according to their distances to ql 0 and ql 1<br />
Step 2: C<strong>al</strong>culate the mean v<strong>al</strong>ues of pixels in each group<br />
to generate these two quantization levels ql 0 and ql 1 .<br />
IJCTA | JULY-AUGUST 2011<br />
Available online@www.ijcta.com<br />
1110
<strong>Aditya</strong> <strong>kumar</strong> <strong>et</strong> <strong>al</strong>, <strong>Int</strong>. J. <strong>Comp</strong>. <strong>Tech</strong>. <strong>Appl</strong>., <strong>Vol</strong> 2 (4), 1099-1117<br />
ISSN:2229-6093<br />
Figure 20(a).bridge512<br />
Figure 20(c).f16-512<br />
Figure 20(e).Pepper512<br />
Figure 20(g).Lena512<br />
Figure 20(i).Baboon512<br />
Figure 20(b).boat512<br />
Figure 20(d).elaine512<br />
Figure 20(f).zelda512<br />
Figure 20(h).goldhill512<br />
Figure 20(j).Barbarra512<br />
Step 3: If either one of the two conditions is held, go to<br />
Step 1.<br />
After the above steps are performed, <strong>al</strong>l pixels in each<br />
block x are classified into G 0 and G 1 and the quantization<br />
levels are stored in ql 0 and ql 1 . The following rule is used to<br />
generate the 1-bit v<strong>al</strong>ue b i in bit map BM for each pixel p i .<br />
b = 0, p ∈ G <br />
1, p ∈ G <br />
Now, the pixels in each block are portioned into two<br />
groups and the quantization levels are c<strong>al</strong>culated. The<br />
compressed codes of one image block x comes out of a tripl<strong>et</strong><br />
(ql 0 , ql 1 , BM).<br />
a) Experiment<strong>al</strong> Results<br />
Some experiment<strong>al</strong> results are presented in this section to<br />
verify the performance of the FBTC <strong>al</strong>gorithm.<br />
Results in Tables V-VII present the comparative<br />
performance of the tradition<strong>al</strong> AMBTC [71] scheme, the<br />
optim<strong>al</strong> grouping AMBTC scheme (OBTC) [49], the<br />
economic<strong>al</strong> AMBTC scheme (EBTC) [64], and the FBTC<br />
scheme when the block sizes are 4×4, 8×8, and 16×16,<br />
respectively.<br />
According to the experiment<strong>al</strong> results in Table V, our<br />
FBTC has considerably reduced the execution time. An<br />
average execution time of 0.033 seconds is consumed in<br />
FBTC scheme [15] when the block size equ<strong>al</strong> to 4×4. It is<br />
about 41.25% of the Performance of proposed scheme is<br />
further improved when the block size is 8×8. The consumed<br />
execution time of the FBTC [49] scheme is 38.7% and 5.4%<br />
of the execution times in Economic<strong>al</strong> BTC and OBTC,<br />
respectively.<br />
Table V. <strong>Comp</strong>arison of FBTC with others techniques when block size is<br />
4×4<br />
Images\Schemes AMBTC OBTC EBTC FBTC<br />
Airplane PSNR 33.253 34.044 33.981 33.838<br />
Time 0.017 0.233 0.067 0.033<br />
Baboon PSNR 27.754 28.199 28.137 28.056<br />
Time 0.017 0.233 0.117 0.033<br />
Barbara PSNR 29.105 29.445 29.384 29.321<br />
Time 0.017 0.250 0.083 0.033<br />
Boat PSNR 31.958 32.540 32.467 32.384<br />
Time 0.017 0.233 0.067 0.033<br />
Lena PSNR 33.643 34.200 34.147 34.020<br />
Time 0.017 0.233 0.067 0.033<br />
Average PSNR 31.143 31.686 31.623 31.524<br />
Time 0.017 0.236 0.080 0.033<br />
Meanwhile, the image qu<strong>al</strong>ity of the proposed scheme is<br />
slightly worse than that of EBTC [12] and OBTC [12] [48].<br />
An average image qu<strong>al</strong>ity gain of 0.537 dB is obtained by<br />
using the proposed scheme compared to the convention<strong>al</strong><br />
AMBTC scheme.<br />
According to the results in table 3.2, the image qu<strong>al</strong>ity of<br />
the FBTC scheme is b<strong>et</strong>ter than that of Economic<strong>al</strong> BTC. In<br />
addition, the proposed scheme requires 35% and 1.8% of the<br />
execution times of EBTC [66] and OBTC [49], respectively.<br />
An average image qu<strong>al</strong>ity gain of 0.655 dB is obtained by<br />
using the proposed scheme compared to the convention<strong>al</strong><br />
IJCTA | JULY-AUGUST 2011<br />
Available online@www.ijcta.com<br />
1111
<strong>Aditya</strong> <strong>kumar</strong> <strong>et</strong> <strong>al</strong>, <strong>Int</strong>. J. <strong>Comp</strong>. <strong>Tech</strong>. <strong>Appl</strong>., <strong>Vol</strong> 2 (4), 1099-1117<br />
ISSN:2229-6093<br />
AMBTC scheme. Based on the results in Tables 3.1 to 3.3, the<br />
improvement of image qu<strong>al</strong>ity increases as the block sizes<br />
increases in the proposed scheme.<br />
Table VI. <strong>Comp</strong>arison of FBTC with others techniques when block size is<br />
8×8<br />
images\schemes AMBTC OBTC EBTC FBTC<br />
Airplane PSNR 30.153 31.036 30.998 30.978<br />
Time 0.017 0.650 0.083 0.033<br />
Baboon PSNR 25.841 26.156 26.127 26.114<br />
Time 0.017 0.683 0.117 0.050<br />
Barbara PSNR 26.653 27.067 27.032 27.025<br />
Time 0.017 0.667 0.100 0.033<br />
Boat PSNR 28.995 29.656 29.609 29.610<br />
Time 0.017 0.633 0.083 0.033<br />
Lena PSNR 30.216 30.883 30.847 30.816<br />
Time 0.017 0.650 0.083 0.033<br />
Average PSNR 28.372 28.960 28.922 28.909<br />
Time 0.017 0.657 0.093 0.036<br />
Table VII. <strong>Comp</strong>arison of FBTC with others techniques when block size<br />
is 16×16<br />
images\schemes AMBTC OBTC EBTC FBTC<br />
Airplane PSNR 27.989 29.025 28.986 28.993<br />
Time 0.017 2.217 0.117 0.030<br />
Baboon PSNR 24.789 25.071 25.057 25.052<br />
Time 0.017 2.383 0.150 0.050<br />
Barbara PSNR 24.837 25.317 25.279 25.291<br />
Time 0.017 2.300 0.150 0.030<br />
Boat PSNR 26.802 27.547 27.508 27.519<br />
Time 0.017 2.233 0.083 0.050<br />
Lena PSNR 27.404 28.279 28.234 28.238<br />
Time 0.017 2.267 0.100 0.050<br />
Average PSNR 26.364 27.048 27.013 27.019<br />
Time 0.017 2.280 0.120 0.042<br />
PSNR than that of YCH can be achieved. Our three stage<br />
compressor gives a bit rate of 0.33 bpp at a PSNR v<strong>al</strong>ue of<br />
33.21 for photos used for ID cards. This compression scheme<br />
may be useful for low cost handheld devices with low<br />
computation<strong>al</strong> power which handles images.<br />
C. Par<strong>al</strong>lel and Element-Reduced Error-Diffused BTC<br />
Block Truncation Coding (BTC) [1] [2] is an efficient<br />
compression technique for its inherent simple coding strategy.<br />
However, the annoying blocking effect and f<strong>al</strong>se contour<br />
accompanied in high coding gain configurations make the<br />
applications relatively limited compares to some up-to-date<br />
compression schemes. For this, in 2010, Jing-Ming Guo is<br />
provided Error-Diffused Block Truncation Coding (EDBTC)<br />
to solve these problems. Whole scenario of this scheme is<br />
shown in Figure 22 and Figure 23.<br />
D. Single Bit Plane based BTC for Color Image <strong>Comp</strong>ression<br />
In 2011, Jun Wang [1] introduced Single bit plane based<br />
BTC for color image compression in LCD overdrive. Single<br />
bit plane is used to represent the R, G and B bit planes and<br />
correct the two representative levels in this scheme. Outcome<br />
of this <strong>al</strong>gorithm can improved the compression ratio of BTC<br />
form 4:1 to 6:1 with the limiting performance losing about<br />
0.297db. This <strong>al</strong>gorithm is provided improved BTC such as<br />
VQ-BTC [1]. Figure 21 is used here for comparative an<strong>al</strong>ysis<br />
of different images with different m<strong>et</strong>hods. PSNR v<strong>al</strong>ue of<br />
each image for different m<strong>et</strong>hods is shown in Figure 21.<br />
B. Low <strong>Comp</strong>utation<strong>al</strong> Scheme based on AMBTC<br />
This scheme is introduced by K.Somasundaram and<br />
I.Kaspar Rajin in 2006 [74] which based on four techniques<br />
which combine used as compilation of current mentioned<br />
<strong>al</strong>gorithm. The four techniques are listed below.<br />
‣ Quad Tree Segmentation<br />
‣ Bit plane Omission<br />
‣ Bit plane coding using 32 visu<strong>al</strong> patterns<br />
‣ <strong>Int</strong>erpolative bit plane coding<br />
The experiment<strong>al</strong> results of this scheme is shown [72] that<br />
this scheme achieve an average bit rate of 0.46 bits per pixel<br />
(bpp) for standard gray sc<strong>al</strong>e images with an average PSNR<br />
v<strong>al</strong>ue of 30.25, which is b<strong>et</strong>ter than the results from the exiting<br />
similar m<strong>et</strong>hods based on BTC.<br />
Low computation<strong>al</strong> complexity is two stage and three stage<br />
gray sc<strong>al</strong>e image compression scheme using quad tree<br />
segmentation. The three stage compressor starts with 16x16<br />
pixel block and the two stage compressor starts with 8x8 pixel<br />
blocks. Experiment<strong>al</strong> results [74], by applying our schemes on<br />
standard images, show that an average bit rate of 0.46 bpp<br />
with an average PSNR v<strong>al</strong>ue of 30.25 is for the three stage<br />
compressor and 0.4923 bpp with an average PSNR v<strong>al</strong>ue of<br />
30.37 for two stage compressor which are b<strong>et</strong>ter in bpp and<br />
Figure 21.PSNR results for different m<strong>et</strong>hods<br />
E. An Improved Algorithm using BSPS and GW<br />
This scheme is introduced by Garima Chopra and A. K.<br />
P<strong>al</strong> in 2011 [4].Geom<strong>et</strong>ric wavel<strong>et</strong> is a recent development in<br />
This <strong>al</strong>gorithm outperformed the EZW, the Bandel<strong>et</strong>s and the<br />
GW <strong>al</strong>gorithm. Current discussing <strong>al</strong>gorithm reported a gain<br />
of 0.22 dB over the GW m<strong>et</strong>hod at the compression ratio [4]<br />
of 64 for the Cameraman test image. Performance an<strong>al</strong>ysis of<br />
this scheme with tradition<strong>al</strong> BTC is tabularized is Table VIII.<br />
Basic<strong>al</strong>ly this table is provided the complexity comparison of<br />
tradition<strong>al</strong> BTC and ODBTC.<br />
Table VIII. <strong>Comp</strong>lexity <strong>Comp</strong>arisons b<strong>et</strong>ween Tradition<strong>al</strong> BTC and ODBTC<br />
(Tot<strong>al</strong> Operations in a Block of Size MXN)<br />
Operations Addition/ Multiplication/division Square root<br />
Subtraction<br />
BTC 2x(mxn)+3 Mxn+9 2<br />
Odbtc Mxn 0 0<br />
IJCTA | JULY-AUGUST 2011<br />
Available online@www.ijcta.com<br />
1112
<strong>Aditya</strong> <strong>kumar</strong> <strong>et</strong> <strong>al</strong>, <strong>Int</strong>. J. <strong>Comp</strong>. <strong>Tech</strong>. <strong>Appl</strong>., <strong>Vol</strong> 2 (4), 1099-1117<br />
ISSN:2229-6093<br />
Figure 22(a).<br />
Figure 22(c). Distribution of the two best diffused directions.<br />
Figure 22(b).<br />
Figure 22(d). Checkerboard pattern and serpentine scan.<br />
Figure 22. Optimization procedure and its fin<strong>al</strong> two diffused elements: (a) optimization procedure (b) element-reduced kernel.<br />
Figure23.<br />
F. VQ based BTC for Color Image <strong>Comp</strong>ression in LCD<br />
Overdrive<br />
This scheme is introduced in 2008 by Jong-Woo Han [3].<br />
This scheme is described a novel block truncation coding<br />
based on the vector quantizer for the color image compression<br />
in LCD overdrive. Due to the constant output bit-rate and the<br />
low computation<strong>al</strong> complexity, this m<strong>et</strong>hod is suitable for the<br />
hardware implementation in LCD overdrive. Experiment<strong>al</strong><br />
results [3] are shown that current m<strong>et</strong>hod achieves higher<br />
compression ratio as well as b<strong>et</strong>ter visu<strong>al</strong> qu<strong>al</strong>ity as compared<br />
with the convention<strong>al</strong> m<strong>et</strong>hods.<br />
G. AM- BTC for Frame Memory Reduction in LCD Overdrive<br />
This scheme is introduced by Jun Wang in 2010 [2]. The<br />
AM-BTC firstly overcomes the limitation by adaptively<br />
selecting 2-level or 4-level BTC according to the edge<br />
property of the coding block. Then, to reduce the bit rate of<br />
AM-BTC he improves the 2-level and 4- level BTCs by using<br />
only luminance bit-map to represent three color bit-maps. As<br />
shown in simulation results [2], the AM-BTC successfully<br />
reduces the frame memory usage to 1/6 and significantly<br />
improved coding performance (up to 3.779 dB) as compared<br />
with other <strong>al</strong>gorithms. When the AM-BTC is applied to LCD<br />
overdrive, it <strong>al</strong>so improves overdrive performance up to 3.390<br />
dB as compared with other comp<strong>et</strong>itive m<strong>et</strong>hod [2].<br />
The whole scenario of this scheme is shown is Figure 24-<br />
26. Figure 24 is shown as the basic idea of <strong>al</strong>gorithm and<br />
figure 24 represents the Block diagram of AM-BTC <strong>al</strong>gorithm.<br />
IJCTA | JULY-AUGUST 2011<br />
Available online@www.ijcta.com<br />
1113
<strong>Aditya</strong> <strong>kumar</strong> <strong>et</strong> <strong>al</strong>, <strong>Int</strong>. J. <strong>Comp</strong>. <strong>Tech</strong>. <strong>Appl</strong>., <strong>Vol</strong> 2 (4), 1099-1117<br />
ISSN:2229-6093<br />
Figure24.Block diagram of overdrive in LCD[2]<br />
Figure25. Block diagram of AM-BTC <strong>al</strong>gorithm [2].<br />
Figure26. Gfig architecture of the AM-BTC encoder<br />
V. COMPARISON TO JPEG<br />
Next we will link tog<strong>et</strong>her the best ideas of the BTC<br />
<strong>al</strong>gorithm. Since the three phases of the BTC were studied<br />
separately, there is no guarantee that the combined <strong>al</strong>gorithm<br />
is optim<strong>al</strong> in any sense, nor the best that could be constructed<br />
from these elements. In spite of that, we will compare it<br />
against JPEG. This is done by first fixing a certain bit rate and<br />
then comparing the MSE-v<strong>al</strong>ues of these two m<strong>et</strong>hods.<br />
The combined BTC <strong>al</strong>gorithm consists of the components<br />
given in Table IX. Hierarchic<strong>al</strong> decomposition is fully<br />
utilized: The corresponding minimum and maximum block<br />
sizes are 2*2 and 32*32. Standard deviation (σ) is used as a<br />
threshold criterion and is s<strong>et</strong> to 6 for other level transitions<br />
than 4*4->2*2, for which the threshold is left as an adjusting<br />
param<strong>et</strong>er.<br />
The bit plane is either stored as such, or its size is reduced<br />
by the interpolation technique. The level of interpolation is<br />
chosen according to the desired bit rate level. It will be<br />
referred by the proportion of coded bits which is 50, 75, or 100<br />
%. Bit planes of uniform blocks are skipped and the 1-bit<br />
quantization is applied to these blocks. The uniform block<br />
skipping threshold is used as a fine-tuning param<strong>et</strong>er.<br />
Figure 26 illustrates the performance of the combined BTC<br />
<strong>al</strong>gorithm for Lena. The tests were performed by varying the<br />
threshold of the hierarchic<strong>al</strong> decomposition from σ=0 to 50.<br />
Each curve represents one of the selected interpolation levels.<br />
In this way we are able to cover reasonably well the bit rates<br />
from 1.0 to 2.0 bits per pixel. B<strong>et</strong>ter image qu<strong>al</strong>ities are still<br />
possible at the cost of decreasing compression efficiency, but<br />
the increase of the bit rate serves the purpose only up to a<br />
limit. By the use of lossless image compression m<strong>et</strong>hods, such<br />
as lossless JPEG [46, 57] or FELICS [25], one can achieve bit<br />
rates of about 4.5 bpp (for Lena). At the other end, the image<br />
qu<strong>al</strong>ity will decrease rapidly if very low bit rates are desired.<br />
A comparison of JPEG and the combined BTC is<br />
summarized in Figure 26-29. Here we use the same param<strong>et</strong>ers<br />
that were used for the test results in Figure 26 with one<br />
exception: we performed two different tests. In the first one,<br />
we used the skipping threshold σ=0, and in the second σ=5.<br />
The fin<strong>al</strong> curves of Figure 26-29 are assembled from the<br />
results of these two so that they correspond to the best MSEv<strong>al</strong>ues<br />
for each bits per pixel selection. The non-continuation<br />
of the curve in Figure 29 originates from the nature of the<br />
image G<strong>al</strong>axy. It consists entirely of low contrast regions, and<br />
therefore the hierarchic<strong>al</strong> decomposition cannot be used for<br />
adjusting its performance as much as is desirable.<br />
In our tests, JPEG clearly outperforms BTC (see Fig. 30).<br />
At the bit rate of 2.0, the MSE of JPEG is 65 % from the MSE<br />
of BTC (for Lena), and at the bit rate of 1.0 bpp, it is only 53<br />
%. The difference in the visu<strong>al</strong> qu<strong>al</strong>ity is y<strong>et</strong> unclear. Sm<strong>al</strong>l<br />
differences in the MSE-v<strong>al</strong>ue are not necessarily reflected in<br />
the visu<strong>al</strong> qu<strong>al</strong>ity and one must keep in mind, that the<br />
distortions caused by these two m<strong>et</strong>hods are not <strong>al</strong>ike.<br />
Moreover, distortion is <strong>al</strong>most unnoticeable if one looks at the<br />
images in their natur<strong>al</strong> sizes. Magnifications of Lena are<br />
shown in Fig. 31. We want to emphasize that there is a definite<br />
need for a b<strong>et</strong>ter objective qu<strong>al</strong>ity measure than MSE.<br />
Table IX<br />
PHASE<br />
METHOD<br />
BLOCK SIZE hierarchic<strong>al</strong> (32→2)<br />
QUANTIZATION SYSTEM gb + means<br />
QUANTIZATION DATA (a,b)<br />
NUBER OF BITS 6+6<br />
coding the quant. data felics<br />
coding the bit plane 1 bpp<br />
interpolation<br />
skipping the bit plane<br />
IJCTA | JULY-AUGUST 2011<br />
Available online@www.ijcta.com<br />
1114
ISSN:2229-6093<br />
<strong>Aditya</strong> <strong>kumar</strong> <strong>et</strong> <strong>al</strong>, <strong>Int</strong>. J. <strong>Comp</strong>. <strong>Tech</strong>. <strong>Appl</strong>., <strong>Vol</strong> 2 (4), 1099-1117<br />
Figure26. Performance of the combined BTC <strong>al</strong>gorithm for Lena<br />
Figure26. <strong>Comp</strong>arison of JPEG versus BTC with test image Lena<br />
Figure28. <strong>Comp</strong>arison of JPEG versus BTC with test image Lena<br />
VI.<br />
CONCLUSION<br />
Fin<strong>al</strong>ly survey literature is concludes as background<br />
information of the variants of the block truncation coding.<br />
Motto of whole literature was that one can remarkably<br />
improve the over<strong>al</strong>l performance of BTC by the use of<br />
hierarchic<strong>al</strong> decomposition and interpolation coding. The most<br />
important change for improving the image qu<strong>al</strong>ity was the use<br />
of 2*2-blocks in the high contrast regions. When compared to<br />
the performance of JPEG, the new combined block truncation<br />
coding is still weak if the image qu<strong>al</strong>ity is measured by the<br />
mean square error.<br />
There are still other aspects which make BTC a useful<br />
<strong>al</strong>ternative for JPEG in some practic<strong>al</strong> applications. All the<br />
selected <strong>al</strong>gorithms are quite simple and easy to implement.<br />
The integration makes the implementation somewhat more<br />
complicated, but it is nevertheless probably much simpler than<br />
the JPEG implementation. Since we did not perform any<br />
throughput an<strong>al</strong>ysis, we can only estimate the running times.<br />
The encoding phase of BTC is very fast, and the decoding<br />
phase is even faster. One can maximize the throughput at the<br />
decoding phase by the use of (a,b) as the quantization data. It<br />
could be further improved by the use of vector quantization<br />
instead of interpolation. This would load the encoding more,<br />
but the computation<strong>al</strong> demands would be much less at the<br />
decoding phase. The hierarchic<strong>al</strong> decomposition has only a<br />
sm<strong>al</strong>l effect on the throughput, but is directly reflected in the<br />
number of blocks to be coded.<br />
By the use of hierarchic<strong>al</strong> decomposition and FELICS<br />
coding, one can reduce the interblock redundancy. This will<br />
eliminate direct access to the decoded file and thus there is no<br />
longer guarantee of fault tolerance.<br />
Fin<strong>al</strong>ly the literature survey on current <strong>al</strong>gorithms which<br />
based on moment preserving <strong>al</strong>gorithm is performed. Each<br />
<strong>al</strong>gorithm which discussed in this literature are tried to<br />
enhance the performance of Block Truncation like <strong>al</strong>gorithm<br />
directly or indirectly. All techniques are provided b<strong>et</strong>ter visu<strong>al</strong><br />
qu<strong>al</strong>ity (PSNR) and good bit rate. Many <strong>al</strong>gorithms like 3.8<br />
and 3.9 are used in LCD overdrive for optimizing the<br />
performance of BTC <strong>al</strong>gorithms. There is MSE, NAE and<br />
WPSNR v<strong>al</strong>ues is <strong>al</strong>so improved.<br />
RFERENCES<br />
[1] Jun Wang, Lin-bo Luo, Kyeong-yuk Min, Yeun-Cheul Jeung, Jong-wha<br />
Chong ,” Single bit plane based block truncation coding for color image<br />
compression in LCD overdrive,” IEEE <strong>Int</strong>ernation<strong>al</strong> Conference on<br />
Consumer Electronics (ICCE), 978-1-4244-4314-7, pp. 53, 2010.<br />
[2] Jun Wang, Jong-Wha Chong ,”Adaptive multi-level block truncation<br />
coding for frame memory reduction in LCD overdrive,” IEEE<br />
Transactions on Consumer Electronics, 0098-3063, pp. 1130, 2011.<br />
[3] Jong-Woo Han, Min-Cheol Hwang, Seong-Gyun Kim, Tae-Ho You,”<br />
Vector quantizer based block truncation coding for color image<br />
compression in LCD overdrive,” IEEE Transactions on Consumer<br />
Electronics, pp. 1839, 0098-3063,2008.<br />
[4] Garima Chopra and A. K. P<strong>al</strong>,” An Improved Image <strong>Comp</strong>ression<br />
Algorithm Using Binary Space Partition Scheme and Geom<strong>et</strong>ric<br />
Wavel<strong>et</strong>s,” IEEE Transactions on Image Processing, vol. 20, no. 1,<br />
January 2011.<br />
[5] Jing-Ming Guo, Member, IEEE, Chih-Yu Lin,” Par<strong>al</strong>lel and Element-<br />
Reduced Error-Diffused Block Truncation Coding,” IEEE Transactions<br />
on Communications, vol. 58, no. 6, June 2010.<br />
[6] David Solomon, “Data <strong>Comp</strong>ression The compl<strong>et</strong>e reference,” 2nd<br />
edition, Springer 2001, Newyork.<br />
[7] E.M. Rafael C. Gonz<strong>al</strong>ez, Richard Eugene; “Digit<strong>al</strong> image processing”.<br />
[8] Bibhas Dhara, Bhabatosh Chanda,” Color image compression based on<br />
block truncation coding using pattern fitting principle,” Journ<strong>al</strong> Pattern<br />
Recognition archive, <strong>Vol</strong>ume 40 Issue 9, September, 2007,Elsevier<br />
Science Inc. New York, NY, USA, Edition 3, page 466, 2008.<br />
[9] T.S. Wang, M.C.Hwang, and S.J. Ko.,” Cost Effective Block Truncation<br />
Coding for Color Image <strong>Comp</strong>ression,” Advanced in Information<br />
Sciences and Service Sciences <strong>Vol</strong>ume 2, Number 3, September 2010.<br />
IJCTA | JULY-AUGUST 2011<br />
Available online@www.ijcta.com<br />
1115
<strong>Aditya</strong> <strong>kumar</strong> <strong>et</strong> <strong>al</strong>, <strong>Int</strong>. J. <strong>Comp</strong>. <strong>Tech</strong>. <strong>Appl</strong>., <strong>Vol</strong> 2 (4), 1099-1117<br />
ISSN:2229-6093<br />
[10] <strong>Aditya</strong> Kumar, Pardeep Singh, “Enhanced Block Truncation Coding for<br />
Gray Sc<strong>al</strong>e Image,” <strong>Int</strong>ernation<strong>al</strong> Journ<strong>al</strong> of <strong>Comp</strong>uter <strong>Tech</strong>nology and<br />
<strong>Appl</strong>ications (2229-6093), <strong>Vol</strong>ume 2 (3), pp. 525-530, 2011..<br />
[11] M. D. Lema and O. R. Mitchell, “Absolute Moment Block Truncation<br />
Coding and Its <strong>Appl</strong>ication to Color Image,” IEEE Transactions on<br />
Communications, <strong>Vol</strong>. 32, pp. 1148-1157, 1984.<br />
[12] <strong>Aditya</strong> Kumar, Pardeep Singh, “Futuristic Algorithm for Gray Sc<strong>al</strong>e<br />
Image based on Enhanced Block Truncation Coding,” <strong>Int</strong>ernation<strong>al</strong><br />
Journ<strong>al</strong> of <strong>Comp</strong>uter Information Systems(2229-5208), <strong>Vol</strong>. 2, No. 2,<br />
pp. 53-59, 2011.<br />
[13] E.M. Rhoma, A.M Abobaker,”Mean square error minimization using<br />
interpolative block truncation coding <strong>al</strong>gorithms ,”2nd <strong>Int</strong>ernation<strong>al</strong><br />
Conference on Education <strong>Tech</strong>nology and <strong>Comp</strong>uter (ICETC), 2010.<br />
[14] Y. C. Hu, "Improved Block Truncation Coding for Image <strong>Comp</strong>ression,"<br />
IEE Electronics L<strong>et</strong>ters, <strong>Vol</strong>. 39, No. 19, pp. 1377-1379, 2003.<br />
[15] M. Rabbani and R. Joshi, “An overview of the JPEG 2000 still image<br />
compression standard,’’ Sign<strong>al</strong> Process. Image Commun. 17, pp. 3–48,<br />
2002.<br />
[16] Pham, T.Q,” Separable bilater<strong>al</strong> filtering for fast video preprocessing’.<br />
Proc. 2005 IEEE <strong>Int</strong>. Conf. on Multimedia and Expo, Amsterdam, the<br />
N<strong>et</strong>herlands, <strong>Vol</strong>. CD1–4, 2005.<br />
[17] Y. C. Hu, "Low-<strong>Comp</strong>lex and Low-bit-rate Image <strong>Comp</strong>ression Scheme<br />
Based on AMBTC," Optic<strong>al</strong> Engineering, <strong>Vol</strong>. 42, No. 7, pp. 1964-<br />
1975, 2003.<br />
[18] C. Y. Yang and J. C. Lin, “EBTC: An Economic<strong>al</strong> M<strong>et</strong>hod for Searching<br />
The Threshold of BTC <strong>Comp</strong>ression,” Electronic L<strong>et</strong>ters, <strong>Vol</strong>. 32, No.<br />
20, pp. 1870-1871, Sep. 1996.<br />
[19] Chih-Chiang Tsou, Shiau-Hua Wu, Yu-Chen Hu,” Fast Pixel Grouping<br />
<strong>Tech</strong>nique for Block Truncation Coding”, proceeding of 2005 work shop<br />
on consumer electronics and sign<strong>al</strong> processing,2005.<br />
[20] Delp E.J., Mitchell O.R., ”Image Coding Using Block Truncation<br />
Coding,” IEEE Transactions on Communications, 27, pp. 1335-1342,<br />
1979.<br />
[21] <strong>Aditya</strong> Kumar, Pardeep Singh, “Aggrandize Bit Plane Coding Using<br />
Gray Code M<strong>et</strong>hod,” <strong>Int</strong>ernation<strong>al</strong> Journ<strong>al</strong> of <strong>Comp</strong>uter <strong>Appl</strong>ications<br />
(0975-8887) <strong>Vol</strong>ume 20– No.6, pp. 44-49, April 2011..<br />
[22] H<strong>al</strong>verson D., Griswold N., Wise G., “A Gener<strong>al</strong>ized Block Truncation<br />
Coding Algorithm for Image <strong>Comp</strong>ression”. IEEE Transactions on<br />
Accoustics, Speech and Sign<strong>al</strong> Processing, 32, pp. 664-668, 1979.<br />
[23] Lema M.D., Mitchell O.R., “Absolute Moment Block Truncation<br />
Coding and its <strong>Appl</strong>ication to Color Images,” IEEE Transactions on<br />
Communications, 32, pp. 1148-1157, 1979.<br />
[24] Efrati N., Liciztin H., Mitchell H.B., “Classified Block Truncation<br />
Coding-Vector Quantization: an Edge Sensitive Image <strong>Comp</strong>ression<br />
Algorithm. Sign<strong>al</strong> Processing, Image Commun. 3, pp. 275-283, 1991.<br />
[25] Goeddel T., Bass S., “A Two Dimension<strong>al</strong> Quantizer for Coding Digit<strong>al</strong><br />
Imagery,” IEEE Transactions on Communications, 29, pp. 60-67, 1989.<br />
[26] Qiu G., Varley M.R., Terrell T.J., “Improved Block Truncation Coding<br />
Using Hopfield Neur<strong>al</strong> N<strong>et</strong>work,” Electronics L<strong>et</strong>ters, 27, pp.1924-1926,<br />
1991.<br />
[27] <strong>Aditya</strong> Kumar, Pardeep Singh, “A Wavel<strong>et</strong> Transform Algorithm for 2 n<br />
Shades Image,” The <strong>Int</strong>ernation<strong>al</strong> Journ<strong>al</strong> on <strong>Comp</strong>uter Science and<br />
Engineering (0975-3397), <strong>Vol</strong>. 3 No. 5, pp. 2136-2144, May 2011..<br />
[28] Gersho A., Gray R.M., “Vector Quantization and Sign<strong>al</strong> <strong>Comp</strong>ression,”<br />
Kluwer Academic Publishers, 1992.<br />
[29] Nasrabadi N.M., King R.A., “Image Coding Using Vector quantization:<br />
A Review,” IEEE Transactions on Communications, 36, pp. 957-971,<br />
1988.<br />
[30] Kunt M., Bénard M., Leonardi R., “Recent results in high-<strong>Comp</strong>ression<br />
Image Coding,” IEEE Transactions on Circuits and Systems, 34, pp.<br />
1306-1336, 1987.<br />
[31] Kunt M., Ikonomopolous A., Kocher M., “Second Generation Image<br />
Coding <strong>Tech</strong>nique,” Proceedings of the IEEE, 73, pp. 549-574, 1985.<br />
[32] Udpikar V., Raina J. “Modified Algorithm for the Block Truncation<br />
Coding of Monochrome Images,” Electronics L<strong>et</strong>ters, 21, pp. 900-902,<br />
1985.<br />
[33] Lloyd S.P., “Least Squares Quantization in PCM,” IEEE Transactions<br />
on Information Theory, 28, pp. 129-137, 1982.<br />
[34] Mitchell H.B., Dorfan M.,” Block Truncation Coding Using Hopfield<br />
Neur<strong>al</strong> N<strong>et</strong>work,” Electronics L<strong>et</strong>ters, 28, pp. 2144-2145, 1992.<br />
[35] Mitchell O.R., Delp E.J., “Multilevel Graphics Representation Using<br />
Block Truncation Coding,” Proceedings of the IEEE, 68, pp. 868-873,<br />
1980.<br />
[36] Wu Y., Coll D.C., “BTC-VQ-DCT Hybrid Coding of Digit<strong>al</strong> Images,”<br />
IEEE Transactions on Communications, 39, pp. 1283-1287, 1991.<br />
[37] Fränti P., Nev<strong>al</strong>ainen O., “Block Truncation Coding with Entropy<br />
Coding,” 1993.<br />
[38] Witten I., Ne<strong>al</strong> R., Cleary J.,”Arithm<strong>et</strong>ic Coding for Data <strong>Comp</strong>ression,”<br />
Comm. ACM, 30, pp. 520-539, 1987.<br />
[39] Tukey J.W., “Exploratory Data An<strong>al</strong>ysis. Addison-Wesley,” 1971.<br />
[40] Weitzman A., Mitchell H.B., “Fast Multi-Stage VQ Search for the BTC-<br />
VQ Algorithm,” Proceedings of 17 th Convention of Electric<strong>al</strong> and<br />
Electronics Engineers in Israel, Tel Aviv, Israel, pp. 182-185, 1991.<br />
[41] Zeng B., “Two <strong>Int</strong>erpolative BTC Image Coding Schemes,” Electronics<br />
L<strong>et</strong>ters, 27, pp. 1126-1128, 1991.<br />
[42] Zeng B., Neuvo Y., “<strong>Int</strong>erpolative BTC Image Coding with Vector<br />
Quantization,” IEEE Transactions on Communications, 41, pp. 1436-<br />
1438, 1993.<br />
[43] Sharma D.J., N<strong>et</strong>rav<strong>al</strong>i A.N., “Design of Quantizer for DPCM Coding of<br />
Picture Sign<strong>al</strong>s,” IEEE Transactions on communications, 25, pp. 1267-<br />
1274, 1977.<br />
[44] Udpikar V., Raina J., “BTC Image Coding Using Vector Quantization,”<br />
IEEE Transactions on communications, 35, pp. 352-356, 1987.<br />
[45] Roy J., Nasrabadi M.,”Hierarchic<strong>al</strong> Block Truncation Coding,” Optic<strong>al</strong><br />
Engineering, 30, pp. 551-556, 1991.<br />
[46] Weitzman A., Mitchell H.B., “An <strong>Int</strong>erblock BTC-VQ Image Coder,”<br />
Proceedings of 11 th IAPR <strong>Int</strong>ernation<strong>al</strong> Conference on Pattern<br />
Recognition, III, Conference C, The Hague, N<strong>et</strong>herlands, pp. 426-429,<br />
1992.<br />
[47] Alcaim A., Oliveira L.V., “Vector Quantization of the Side Information<br />
in BTC Image Coding,” Proc. ICCS/ICITA 92, Singapore, 1, pp. 345-<br />
349, 1992.<br />
[48] Nasiopoulos P., Ward R., Morse D., ”Adaptive <strong>Comp</strong>ression Coding,”<br />
IEEE Transactions on Communications, 39, pp. 1245-1254, 1991.<br />
[49] Chih-Chiang, Tsou, Shiau-Hua Wu, and Yu-Chen Hu,” Fast Pixel<br />
Grouping <strong>Tech</strong>nique for Block Truncation Coding,” workshop on<br />
consumer and sign<strong>al</strong> processing, pp.1-6, 2005.<br />
[50] Arce G., G<strong>al</strong>lagher N.C. Jr., ”BTC Image Coding Using Median Filter<br />
Roots,” IEEE Transactions on Communications, 31, pp. 784-793, 1983.<br />
[51] Wang Q., Neuvo Y., “D<strong>et</strong>erministic Properties of Separable and Cross<br />
Median Filters with an <strong>Appl</strong>ication to Block Truncation Coding,”<br />
Multidimension<strong>al</strong> Systems and Sign<strong>al</strong> Processing, 4, pp. 23-38, 1993.<br />
[52] Efrati N., Liciztin H., Mitchell H.B., “Classified Block Truncation<br />
Coding-Vector Quantization: an Edge Sensitive Image <strong>Comp</strong>ression<br />
Algorithm,” Sign<strong>al</strong> Processing, Image Commun. 3, pp. 275-283, 1991.<br />
[53] Kamel M., Sun C. Guan L. “Image <strong>Comp</strong>ression by Variable Block<br />
Truncation Coding with Optim<strong>al</strong> Threshold,” IEEE Transactions on<br />
Sign<strong>al</strong> Processing, 39, pp. 208-212, 1991.<br />
[54] <strong>Aditya</strong> Kumar, Pardeep Singh, “<strong>Comp</strong>arative study of Block Truncation<br />
Coding with EBTC,” The <strong>Int</strong>ernation<strong>al</strong> Journ<strong>al</strong> of Advances in<br />
<strong>Comp</strong>uter N<strong>et</strong>works and Security, 2011.( In Press).<br />
[55] Rabbani M., Jones P.W., “Digit<strong>al</strong> Image <strong>Comp</strong>ression <strong>Tech</strong>niques,”<br />
Bellingham, USA, SPIE Optic<strong>al</strong> Engineering Press, 1991.<br />
[56] Delp E.J., Mitchell O.R., “The Use of Block Truncation Coding in<br />
DPCM Image Coding,” IEEE Transactions on Sign<strong>al</strong> Processing, 39, pp.<br />
967-971, 1991.<br />
[57] Mitchell O.R., Bass S.C., Delp E.J., Goeddel T.W., Huang T.S., ”Image<br />
Coding for Photo An<strong>al</strong>ysis,” Proc. Soc. Inform. Display, 21, pp. 279-<br />
292, 1980.<br />
[58] Mon<strong>et</strong> P., Dubois E., “Block Adaptive Quantization of Images,” IEEE<br />
Transactions on Communications, 41, pp. 303-306, 1993.<br />
IJCTA | JULY-AUGUST 2011<br />
Available online@www.ijcta.com<br />
1116
<strong>Aditya</strong> <strong>kumar</strong> <strong>et</strong> <strong>al</strong>, <strong>Int</strong>. J. <strong>Comp</strong>. <strong>Tech</strong>. <strong>Appl</strong>., <strong>Vol</strong> 2 (4), 1099-1117<br />
ISSN:2229-6093<br />
[59] DeWitte J., Ronsin J., “Origin<strong>al</strong> Block coding Scheme for Low Bit Rate<br />
Image Transmission,” Sign<strong>al</strong> Processing II: Theories and <strong>Appl</strong>ications,<br />
Proc. EUSIPCO-83. Elsevier, Amsterdam, pp. 143-146, 1983.<br />
[60] Wang Y., Mitra S.K., “Image Representation Using Block Pattern<br />
Models and Its Image Processing <strong>Appl</strong>ications,” IEEE Transactions on<br />
Pattern An<strong>al</strong>ysis and Machine <strong>Int</strong>elligence, 15, pp. 321-336, 1993.<br />
[61] Cheng S.-C., Tsai W.-H., “Image <strong>Comp</strong>ression Using Adaptive<br />
Multilevel Block Truncation Coding,” Journ<strong>al</strong> of Visu<strong>al</strong> Communication<br />
and Image Representation, 4, pp. 225-241, 1993.<br />
[62] Eskicioglu A.M., Fisher P.S., “A Survey of Qu<strong>al</strong>ity Measures for Gray<br />
Sc<strong>al</strong>e Image <strong>Comp</strong>ression,” Proceedings Space and Earth Science Data<br />
<strong>Comp</strong>ression Workshop, Snowbird, Utah, pp. 49-61, 1993.<br />
[63] Uhl T.J., “Adaptive Picture Data <strong>Comp</strong>ression Using Block Truncation<br />
Coding,” Sign<strong>al</strong> Processing II: Theories and <strong>Appl</strong>ications, Proc.<br />
EUSIPCO-83. Elsevier, Amsterdam, pp. 147-150, 1983.<br />
[64] P. Franti, P. Nev<strong>al</strong>ainen, and T. Kaudoranta, "<strong>Comp</strong>ression of Digit<strong>al</strong><br />
Image by Block Truncation Coding: A Survey," <strong>Comp</strong>uter Journ<strong>al</strong>, <strong>Vol</strong>.<br />
37, No. 4, pp. 308-332, 1994.<br />
[65] Y. G. Wu and S. C. Tsai, "An Efficient BTC Image <strong>Comp</strong>ression<br />
<strong>Tech</strong>nique," IEEE Trans. On Consumer Electronics, <strong>Vol</strong>. 44, No. 2, pp.<br />
317-325, 1998.<br />
[66] Y. C. Hu, "Low-<strong>Comp</strong>lex and Low-bit-rate Image <strong>Comp</strong>ression Scheme<br />
Based on AMBTC," Optic<strong>al</strong> Engineering, <strong>Vol</strong>. 42, No. 7, pp. 1964-<br />
1975, 2003.<br />
[67] C. Y. Yang and J. C. Lin, “EBTC: An Economic<strong>al</strong> M<strong>et</strong>hod for Searching<br />
The Threshold of BTC <strong>Comp</strong>ression,” Electronic L<strong>et</strong>ters, <strong>Vol</strong>. 32, No.<br />
20, pp. 1870-1871, Sep. 1996.<br />
[68] Lu W.W., Gough M., Davies P., “Scientific Data <strong>Comp</strong>ression for<br />
Space: a Modified Block Truncation Coding Algorithm,” Proc. SPIE,<br />
1470, pp.197-205, 1991.<br />
[69] Wu Y., Coll D.C., “Single Bit-Map Block Truncation Coding of Color<br />
Images,” IEEE Journ<strong>al</strong> on Selected Areas in Communications, 10, pp.<br />
952-959, 1992.<br />
[70] He<strong>al</strong>y D., Mitchell O.R.,”Digit<strong>al</strong> Bandwidth <strong>Comp</strong>ression Using Block<br />
Truncation Coding,” IEEE Transactions on Communications, 29,<br />
pp.1809-1817, 1981.<br />
[71] Griswold N., H<strong>al</strong>verson D., Wise G., “A Note on Adaptive Block<br />
Truncation Coding for Image Processing,” IEEE Transactions on<br />
Accoustics, Speech and Sign<strong>al</strong> Processing, 35, pp. 1201-1203, 1987.<br />
[72] K.Somasundaram and I.Kaspar Raj, “An Image compression Scheme<br />
based on Predictive and <strong>Int</strong>erpolative Absolute Moment Block<br />
Truncation Coding”, GVIP Journ<strong>al</strong>, <strong>Vol</strong>ume 6, Issue 4, December, 2006.<br />
[73] <strong>Aditya</strong> Kumar, Pardeep Singh, “An Image <strong>Comp</strong>ression Algorithm for<br />
Gray Sc<strong>al</strong>e Image,” IEEE <strong>Int</strong>ernation<strong>al</strong> Conference on Emerging Trends<br />
in N<strong>et</strong>works and <strong>Comp</strong>uter Communications, 978-1-4577-0238-9,<br />
pp.342-346, 2011.<br />
[74] K.Somasundaram,I.KasparRaj,“Low <strong>Comp</strong>utation<strong>al</strong> Image <strong>Comp</strong>ression<br />
Scheme based on Absolute Moment Block Truncation Coding,” World<br />
Academy of Science, Engineering and <strong>Tech</strong>nology 19 2006.<br />
AUTHORS PROFILE<br />
<strong>Aditya</strong> Kumar received his B. <strong>Tech</strong>. degree<br />
from Ajay Kumar Garg Engineering College,<br />
Ghaziabad. He is currently pursuing M. <strong>Tech</strong>. at the<br />
Department of <strong>Comp</strong>uter Science & Engineering,<br />
Nation<strong>al</strong> Institute of <strong>Tech</strong>nology, Hamirpur (HP)<br />
India. His research interests include Image<br />
compression and Digit<strong>al</strong> Image processings. He is<br />
currently working on Image <strong>Comp</strong>ression<br />
<strong>Tech</strong>niques.<br />
Pardeep Singh is working as Assistant<br />
Professor in the department of <strong>Comp</strong>uter Science &<br />
Engineering of Nation<strong>al</strong> Institute of <strong>Tech</strong>nology,<br />
Hamirpur (HP) India. He is perusing PhD from the<br />
parent institute. His research area is speech<br />
processing. He is a member of <strong>Int</strong>ernation<strong>al</strong><br />
Association of <strong>Comp</strong>uter Science and Information<br />
<strong>Tech</strong>nology (IACSIT) and <strong>Int</strong>ernation<strong>al</strong> Association<br />
of Engineers (IAENG).<br />
IJCTA | JULY-AUGUST 2011<br />
Available online@www.ijcta.com<br />
1117