form. One of the issues to be studied is precisely at what point in the overall clearing process to capture the check image and thus truncate the pape

Transcription

1 Check Image Compression Using A Layered Coding Method Jincheng Huang y, Yao Wang y, and Edward K. Wong z y Department of Electrical Engineering Polytechnic University, Brooklyn, New York z Department of Computer and Information Science Polytechnic University, Brooklyn, New York, ABSTRACT An emerging trend in the banking industry is to digitize checks for storage and transmission. An immediate requirement for ecient storage and transmission is check image compression. General purpose compression algorithms such as JPEG and wavelet-based methods produce annoying ringing or blocking artifacts at high compression ratios. In this paper, a layered approach tocheck image compression is proposed, with which acheck image is represented in several layers. The rst layer describes the foreground map; the second layer species the graylevels of foreground pixels; the third layer is a lossy representation of the background image; and the fourth layer describes the error between the original and the reconstructed image of the rst three layers. The layered coding approach produces images of better quality than traditional JPEG and wavelet coding methods, especially in the foreground, i.e., the text and graphics. In addition, this approach allows progressive retrieval or transmission of dierent image layers. Keywords: Layered Coding, Image Compression, Image Segmentation, Check Image Processing 1. INTRODUCTION The Financial Services Technology Consortium (FSTC) has recently completed a two-year Interbank Check Imaging (ICI) project, launched to nd a better way of managing the more than 63 billion checks processed by banks and nancial institutions each year. Polytechnic University participated as an advisory member of the ICI project. One of the goals of the project was to demonstrate that digital imaging allows truncation of paper checks at earlier points of the clearing process. Check truncation can cut the total cost of processing each check, leading to substantial savings to the U.S. banking industry, and ultimately, to the customer. One major accomplishment of the ICI project was the incorporation of the newly adopted industry standard for image interchange (ANSI X9.46) [1] into a pilot test environment. The new X9.46 Check Processing standard is a exible specication designed to meet diverse nancial applications. It supports interchange of digitized check images between nancial institutions, and is useful for a variety of banking applications such as image archive, image-based return item notication, image-based signature verication, forward presentment of nancial data, corporate positive pay services, and others. The proposed check truncation process would eliminate the paper check from the clearing process, so that only a statement entry is delivered back to the originator. However, the original paper check would be retained in image 1

2 form. One of the issues to be studied is precisely at what point in the overall clearing process to capture the check image and thus truncate the paper. Here, dierent compression techniques may have dierent impacts on the overall architectural design of the proposed system. In particular, the amount ofcheck image compression will aect the bandwidth requirement in the interchange process, and the storage requirement in archiving them. With the 63 billions checks processed in the United States per year, and a storage period of seven years on all returned checks as required by federal laws, any amount of improvement in compression will lead to signicant savings. A standard personal check measures approximately 36 square inches (front and back). Corporate checks can vary widely from about 20 to 60 square inches. A simple approach toscanning with 100 dpi (dot per inch) and 8-bit grayscale can lead to an image with 360 Kbytes of information per check. Commercial systems now in use employ standard compression techniques like JPEG to reduce the information to about 42 Kbytes per check. Other general purpose image compression methods such as the wavelet method could also be used. At high compress ratios, the JPEG and the wavelet codecs produce annoying artifacts near the printed and written characters and symbols. Moreover, the JPEG method yields blocking and severe blurring eects at high compression ratios. Twotypical check images and their compressed versions using the JPEG and a wavelet coder are shown in Figures 1 to 6, respectively. We can see the aforementioned compression artifacts easily. By examining the images in Figures 1 and 4, we see that a check image can be viewed to compose of a foreground overlaid on a background. The foreground includes written or printed numerals and text, line segments, and logo. It is usually homogeneous in gray level value and appears dark in color. On the other hand, the background usually contains smoothly-varying scenery, or light-intensity regular patterns or textures. In view of these characteristics, we developed a layered coding method for check image compression. In this approach,acheck image is separated into four dierent layers: the binary map of the foreground region, the gray-scale values of foreground pixels, a compressed version of the background image, and the residual image representing the error between the original image and the reconstructed image from the rst three layers. Since these layers have dierent characteristics, a dierent compression algorithm can be employed for each layer independently. Figure 7 shows the block diagram of the layered coding approach. The main contribution of this paper is the idea of splitting the check information into layers to allow image regions with dierent characteristics to be treated dierently, and to allow progressive information retrieval. For compression of the foreground and background layers, we choose to use state-of-the-art algorithms for coding binary images and grayscale images, respectively. A key to obtaining high coding eciency within this layered coding framework is an eective segmentation algorithm for separating the foreground from the background. We have developed a segmentation algorithm that uses an adaptive morphological lter, which varies the size of the structuring element based on the estimated width of the line patterns in the foreground. In the remainder of this paper, we rst present our segmentation algorithm in Section 2. Then the coding algorithm for dierent layers are described in Section 3. Computer simulation results are presented in Section 4. Section 5 concludes the paper. 2

3 2. SEGMENTATION OF FOREGROUND FROM BACKGROUND A crucial step in the layered approach is to segment an image into the foreground and the background using a morphological approach. In [2], binary morphological operations were used to segment straight line segments from a binary document image. In our problem, segmentation is performed in the grayscale domain using grayscale morphological operation. In this section, we rst review basic denitions of mathematical morphology operations in section More details on this subject can be found in [3{6]. We then describe our segmentation algorithm Mathematical Morphology Mathematical morphology is based on set theory. There are two basic morphological operations: dilation and erosion. With A and B as sets in the 2-D integer space Z 2, the dilation ofaby B, denoted by A B, is dened as A B = fxj( ^B)x \ A 6= ;g; (1) where B is commonly referred to as the structuring element, ^B is the reection of B, which is dened as ^B = fxjx =,b; for b 2 Bg, and ; denotes the empty set. The erosion ofaby B, denoted A The opening ofaby B, denoted A B, is dened as Finally the closing ofaby B, denoted A B, is dened as B, is dened as A B = fxj(b) x Ag: (2) A B =(A B) B: (3) A B =(A B) B: (4) Opening has the eect of smoothing the contours of an image, breaks narrow isthmuses, and eliminates thin protrusions. A closing operation smooths sections of contours, fuses narrow breaks and long thin gulfs, eliminates small holes, and lls gaps in the contour. Grayscale morphological image operations are extensions of binary morphological operations into the grayscale domain. Let f(x; y) represents a grayscale image function, and b(x; y) a grayscale structuring element. Grayscale dilation of f by b is dened as (f b)(x; y) =maxff(x, m; y, n)+b(m; n)j(x, m); (y, n) 2 D f ;(m; n) 2 D b g; (5) where D f and D b are the domains of f and b, respectively. Grayscale erosion of f by b is dened as (f b)(x; y) =minff(x + m; y + n), b(m; n)j(x + m); (y + n) 2 D f ;(m; n) 2 D b g: (6) The opening of image f by b is f b =(f b) b: (7) 3

4 The closing of image f by b is f b =(f b) b: (8) The opening of f by b may beinterpreted geometrically as the process of pushing the structuring element against the underside of the surface as shown in Figure 8. All the valleys that are narrow with respect to the structuring element are reduced in sharpness. Figure 9 shows the geometrical interpretation of closing in a one-dimensional signal. The structuring element slides on top of the surface. The narrow valleys are lled in while peaks are left in the original form Segmentation of Background and Foreground When the intensity values of the foreground and background are widely separated in a check image, e.g., the image in Figure 1, a simple thresholding operation can be used to separate the foreground from the background. To handle textual images with background shading, Mitchell et al. [7] proposed a more advanced thresholding technique. However, for check images with texture or scenery background that varies in intensity values, it is very dicult to separate the background and foreground by such thresholding methods. The basic segmentation operation used here is morphological closing. Figure 10 shows the basic block diagram of our foreground/background segmentation algorithm. A closing operation is rst applied to the input image. The original image is then subtracted from the resulting image. The residual image is then thresholded to generate a binary foreground map. We then use an area-ratio detection algorithm [8] to eliminate isolated pixels or unwanted short branches in the binarized image. This algorithm changes a foreground pixel to a background pixel if the percentage of foreground pixels in a window, which centers at the foreground pixel, is less than a specic threshold. In our previous study [8], a morphological grayscale transform with a cone-shaped structuring element was used for check images digitized at 100 dpi. Akey to the success of the algorithm is that the diameter of the cone must match closely the width of the foreground lines to be extracted. For test images digitized with varying resolutions, the width of the thin lines wewant to extract can vary from a few pixels to fty pixels. Small structuring elements fail to extract thick lines and large features such as bank logos. A large structuring element is required to extract large features. However, the use of large structuring elements over the entire image will increase computation time and may also falsely extract some unwanted artifacts on checks with a rich background. For detecting lines and features of variable size, a variable size structuring element should be used. Toward this goal, we have developed a two-pass algorithm. The input image is rst divided into small blocks of size N M. In the rst pass, we determine the size of the structuring element to be used in each small block; and in the second pass, we perform closing operation using the determined structuring element size, followed by a thresholding operation over each block. As we can see from Figure 9, closing with a structuring element of the right shape and size can smooth the abrupt transition in a signal. Note that the boundary of a valley with steep transition can be smoothed even if the valley is wider than the structuring element. In the rst pass, to determine the appropriate structuring element size, we 4

5 need to nd the diameter of the valleys in the current block ifthey exist. To detect the valley boundary, we use the property that the transition from background to foreground is abrupt. Therefore, the boundary of potential foreground regions can be detected by grayscale closing with a small cone-shaped structuring element, followed by subtraction of the original image and a thresholding operation. The size of the structuring element for each block is then determined by applying a scanning algorithm to the resulting binarized image. The algorithm scans the block in the horizontal, vertical and two diagonal directions. We assume that there may beseveral separated valleys along each scan line in a given direction. The algorithm scans the block line by line and determines the width of each valley formed and then determines the maximum width. Figure 11 shows the scanning procedure used to determine the maximum width of valleys in one given direction. The size of the structuring element inablockistaken to be the smallest of the maximum widths scanned in the four directions. In the second pass, each block is expanded by half the size of the structuring element along the four boundaries, so the closing operation can be applied to the entire block inside the original boundary. After the closing operation, data outside the original block boundary are discarded. When the closing operation is completed for all the blocks, the original image is subtracted from the resulting image. Thresholding is applied to the residual image using a global threshold. An area-ratio algorithm is then applied to the binarized image to remove isolated pixels. Results obtained after the rst pass and the second pass for the images shown in Figures 1 and 4 are presented in Figures 13 to 16. We can see that for the check image in Figure 1, which has only thin lines, the rst pass alone is sucient. But for the check image in Figure 4, the foreground features vary in size and are hard to extract with a xed size structuring element. The proposed two-pass algorithm proves to be quite eective for this type of images. 3. THE LAYERED CODING ALGORITHM As shown in Figure 7, the layered coding algorithm rst performs background/foreground segmentation on the input image. Two outputs are generated for the foreground: the foreground map (layer I) is a binary image that masks all the foreground pixel locations; the foreground grayscale image (layer II) consists of the original grayscale values of all foreground pixels. The background image (layer III) is obtained by removing the foreground pixels from the input image. Finally, the error image (layer IV) is the dierence between the original image and the reconstructed image from the rst three layers. Since each layer has its distinct characteristics, dierent compression methods are used for dierent layers. In this section, we describe the coding algorithm for each layer First Layer Coding The rst layer, which is the binary foreground map, is coded by the JBIG method [9,10]. To optimize the coding gain, the sequential mode is used. This layer is for the purpose of masking the foreground pixels. This layer alone is sucient when only a binary check image is needed. For example, the foreground map is all that is necessary for printing images of returned checks to customers or for general browsing of a check image database. 5

6 3.2. Second Layer Coding The second layer, which consists of the gray scale values of foreground pixels, is coded by block-based DPCM. In this approach, the entire image is segmented into 8x8 blocks, and if a block includes any foreground pixels, the average of those foreground pixels is calculated. Then the average value of the previous foreground block is subtracted from the current average. The dierence value is then compressed by Human coding. Because the gray level variation of the foreground is very small (as printed or written letters usually have a close to uniform intensity, and hence the dierence values are usually small), the dierence values have avery narrow distribution and the Human coder can achieve a high compression ratio. To achieve even higher compression gains, the arithmetic coding method could be applied. When the second layer image is being reconstructed, every foreground pixel in the block is lled with the same intensity value. The addition of this layer gives more truthful representation of the foreground than the rst layer alone. For checks with high dollar amounts, or for research into questionable checks by the banks, the gray level information of the foreground is useful Third Layer Coding The third layer is the background image after removing the foreground pixels. It could be an image of smoothly varying scene or of regular repetitive texture. Note that the gaps originally occupied by foreground pixels must be lled in properly. If these gaps are replaced with a constant level, say black or white, articial high frequency contents will be created in the background image, thus lowering the compression ratio. To improve the compression eciency in the third layer, the removed foreground pixels are replaced by the average of the neighboring background pixel values. In the decoder, these pixels are set to the foreground gray levels specied in layer II. Examples of the original background images and those after smoothing are shown in Figures 17 to 20. The quality and compression ratio of the overall image is largely determined by this stage, as the rst two layers are represented nearly losslessly and the required number of bytes is relatively constant and is determined by the complexity of the foreground. We have explored the use of JPEG [11] and the set partitioning in hierarchical trees wavelet (SPIHT) coder by Said and Pearlman[12,13]. We have found that the JPEG method is slightly better with background images consisting of graphics textures, as in Figure 1, while the wavelet coder is more suitable for background images composed of natural scenery, as in Figure 4. By varying the quantization parameter in the JPEG coder or the desired le size in the SPIHT coder, we can arrive at a wide range of image quality and associated compression ratios. We have found through experiments that by requiring the size of the third layer to be slightly higher than of the rst two layers, the reconstructed image is usually visually acceptable Fourth Layer Coding The fourth layer is the dierence between the original and the reconstructed image obtained by combining the rst three layers. If a perfect reconstruction is required, this layer can be coded by a lossless compression method such as arithmetic or Human coding. The coding for this layer is not considered in our present study. 6

7 4. SIMULATION RESULTS To evaluate the performance of the above layered coding algorithm, we have collected a total of 13 dierent personal checks, and the front side of each is digitized with spatial resolutions of 100 dpi, 200 dpi, and 300 dpi, respectively. We evaluated the performance of two versions of the proposed layered coding method: one uses the JPEG coder (using the PVRG-JPEG Code [14]) to code layer III (denoted as LCj) and the other one uses the SPIHT (using the code from [13]) method (denoted as LCw). For comparison purposes, we also applied the JPEG method (using the PVRG-JPEG Code [14]) and the SPIHT coder (using the code from [13]) to the original images directly. In the layered coding implementation, we used a cone-shaped structuring element with a radius of 4 pixels in the rst-pass to outline the potential foreground areas. In the second pass, a variable size, disc-shaped structuring elements is used. To generate the Human table for the second layer, we selected 21 check images in the training set with 7 images from each digitization resolution. The dynamic range of the dierence values is 255. However, the large values rarely occur. Hence, we only included a limited number ofvalues and an Escape codeword in the Human table. For those values which are not included in the table, the Escape codeword is rst specied and the original value (with xed length of 8 bits) follows. We have found that 69.6% of all values are in the range of 15, 97.4% are in the range of 63 and 99.7% are in the range of 95. We examined three cases for Human coding: the rst one includes values in the range of 15 (referred as Range15); the second one uses a range of 63 (referred as Range63) and the last one uses a range of 95 (referred as Range95). Theoretically, the more values we include in the Human table, the greater compression gain we can get. The Human coding results vs. the original data amount isshown in Figure 12. The performance obtained with Range63 improves signicantly over that of Range15. However, the performance of Range95 is only about 1% better than Range63. The improvement margin diminishes as more values are included. Hence, we use Range95 for Layer II coding. Figures 13 to 26 show the experimental results of the layered coding algorithm for the check images previously shown in Figures 1 and 4. The reconstructed images of layer I alone are shown in Figures 14 and 16. The reconstructed images from Layers I and II are shown in Figures 21 and 24. The reconstructed images by LCj and LCw are shown in Figures 22, 23, 25 and 26. These images should be compared to those obtained by JPEG and SPIHT, previously shown in Figures 2, 3, 5, and 6. The three algorithms are congured such that they achieve about similar compression ratios (CR = 20 for 100 dpi images, CR = 50 for 200 dpi images) so that we can compare the resulting image quality. It is easy to see the ringing eect near the numerals and characters in the JPEG (Figures 2 and 5) and SPIHT (Figures 3 and 6) compressed images. The blocking eect in the JPEG compressed image (Figures 2 and 5) is also highly noticeable. The background of the SPIHT compressed image shows blurring eect. The layered coding method can preserve the sharpness of the text and avoid the ringing artifact. Even for the background, the layered coding method has better visual quality than when applying the JPEG or wavelet codec to the entire image because the background image in the layered coding method is more homogeneous than the original image. Figures 27 to 7

8 30 show the images compressed by JPEG and SPIHT at lower compression ratios, but have similar visual quality in the backgrounds as the images compressed by the layered coding (Figures 22, 23, 25, and 26). Even though the background regions are rendered better, we can still see the ringing artifacts surrounding the texts in the foreground. This artifact is more pronounced in the JPEG compressed lower resolution image (Figure 27). We found that the SPIHT method has better performance than the JPEG method on smoothly-varying type of background. However, it does not perform well for thin-line type of patterns such as those in the check image shown in Figure 1. The SPIHT method has better performance for images digitized at high resolutions because pixel values change more gradually in these images. As can be seen from the above examples, layer I alone provides necessary information for many banking applications, while layers I and II together give a more truthful representation of the foreground. Table 1 shows the average bit allocation for each layer in our experiment using the layered approach with JPEG for dierent spatial resolutions. The bit allocations for the layered coding with SPIHT are similar. As we can see from these tables, about 30% and 10% of total bits are allocated to layers I and II, respectively. Table 2 summarizes the average total byte counts and compression ratios for dierent layer combinations. If only layer I is required in a given banking application, the compression ratio is up to 65 for images digitized at 100 dpi; about 115 for images digitized at 200 dpi; and about 155 for images digitized at 300 dpi. On the other hand, if both layer I and II are required, the compression ratios are 50, 80, and 110, respectively. Once the foreground binary map is generated, the byte count for the foreground (layer I and layer II) is relatively xed. The third layer can be adjusted for dierent visual quality. The overall compression ratio for the rst three layers depends signicantly on the coding of layer III. Table 1 shows that about 60% of the total bits are used in layer III for a visually acceptable quality. For a quantitative comparison, Figures 31 to 33 compare the PSNR and compression ratios obtained by dierent algorithms at various spatial resolutions. The visual image quality for the results quoted here are similar to those shown in Figures 22 and 25 (for LCj), 23 and 26 (for LCw), 27 and 29 (for JPEG), and 28 and 30 (for SPIHT). They have similar quality in the background, but the foreground letters and line graphics are rendered better with the LCj and LCw methods. It can be seen that the PSNR for our proposed algorithm is about 2{3 db lower than JPEG or about 3{4 db lower than SPIHT. This is mainly due to the large error in the lossy representation of the foreground gray levels, although these errors are hardly visible to our eyes. For similar visual quality at the background, the LCj and LCw methods achieve higher compression ratios. As already shown in previous gures, although the PSNR is lower with the LCj and LCw methods, the visual quality is substantially better than JPEG and SPIHT. Table 3 lists the average compression ratios for all test images, obtained at dierent spatial resolutions. At similar visual quality for the background, the compression ratio for the layered coding method is about 50% to 100% higher than those of JPEG or SPIHT at low resolution (100 to 200 dpi). As resolution increases, the gain decreases. But still the foreground information is sharper with the layered coding method. As for the two layered coding methods, LCw achieves better results than LCj on the average. 8

9 5. DISCUSSION Compression ratio is a key parameter when comparing the performance of dierent check image compression algorithms, as it directly aects the storage requirement for archiving check images, and the bandwidth requirement for transmitting check images. However, comparison of the compression ratios obtained with dierent compression algorithms is meaningful only when image quality is considered at the same time. It is obvious that higher compression ratios can always be obtained at the expense of increased loss of detailed information, thus decreased image quality. Here, for banking applications, the image quality measure used is tied to the requirement of specic banking functions. For signature verication, the quality measure used could be related to the sharpness or legibility ofthe signature; whereas for general viewing, the quality measure used is related to the overall perceived quality ofthe check image. Opinions of bank experts are needed here to come up with these quality measures. We have developed and tested a layered coding method on personal bank checks digitized at dierent resolutions. The layered coding method has outperformed the JPEG and SPIHT methods in terms of visual quality for check images compressed at about the same compression ratios. The layered coding method preserves the sharpness and legibility of the foreground information more accurately and is free from the ringing artifacts associated with the JPEG/wavelet coding method. The foreground contains important information for many banking applications. As the proposed method separates check images into separate layers, it also facilitates progressive or independent retrieval or transmission of the stored image layers for dierent banking requirements and functions. For many banking applications, such as general browsing or printing of returned checks for customers, the rst layer (binary foreground map) is sucient. For certain applications, the second layer may also be necessary. For more complete and truthful representation (such as archives), the rst three layers may bedesired. Finally, for special checks (with large courtesy amount or with special security requirement), the fourth layer may be required. The main contribution of this paper is the proposed layered coding framework for compression of check images. Our main objective istoshow the improvement of such alayered scheme over a generic image coder that processes the entire image in a single layer, e.g., JPEG or SPIHT. A key to the success of this coding framework is the eectiveness of the segmentation algorithm for separating foreground from background. The background patterns in check images can vary from a plain background to a repeated texture pattern, and from natural scenery to complex graphics. The segmentation algorithm we have developed can work quite well in a variety ofcheck images we have tested. But it can still fail in certain cases. As a future work, more eective segmentation algorithms may be developed, which will contribute to the improvement of the coding eciency as well as the information content of the rst layer. For the compression of individual layers, methods other than JBIG, DPCM, JPEG or SPIHT may bedeveloped, to better match the statistics of check images. We propose to use these algorithms because the hardware/software for these methods are readily available and their performances are comparable to other alternatives. 9

10 6. Acknowledgments The authors gratefully acknowledge the discussions they had with Professor Sal Stolfo of Columbia University, onthe use of a model-based or background subtraction technique for check image compression. In this approach, images of model blank checks are stored and subtracted from an input check image. The dierence image can then be eciently coded. These discussions motivated and led the authors in the development of the layered-approach described in this paper. The authors would like to thank the nancial support provided by the New York State Center for Advanced Technology in Telecommunications (CATT) to work on this project, and the nancial support provided by the Financial Services Technology Consortium (FSTC) for participation in the Interbank Check Imaging (ICI) project. REFERENCES 1. X9/TG-15, Technical Guideline for the Use and Understanding of ANSI X9.46, Draft, American National Standards Institute (ANSI), E. K. Wong, Y. Chen, and S. Y. Ohn, \Straight Line Detection in Binary Drawing Using Morphological Operation," Proc. Nonlinear Image Processing IX, IS&T/SPIE, 10 th Annual Symposium on Electronic Imaging, San Jose, CA, Jan 24-30, J. Serra, Image Analysis and Mathematical Morphology, Academic Press, New York, C. R. Giardina, and E. R. Dougherty, Morphological Methods in Image and Signal Processing, Prentice-Hall, Englewood Clis, N.J., E. H. Liang and E. K. Wong, \Hierarchical Algorithms for Morphological Image Processing," Pattern Recognition, pp. 511{529, April R. M. Haralick, S. R. Sternberg, and X. Zhuang, \Image Analysis using Mathematical Morphology," IEEE Trans. Pattern Anal. and Mach. Intell., PAMI-9, pp , J. L. Mitchell, W. B. Pennebaker, D. Anastassiou, and K. S. Pennington, \Graphics Image Coding for Freeze- Frame Videoconference," IEEE Transactions on Communications, Vol. 37, No. 5, pp. 515{522, May A. Susanto, Y. Wang, and E. K. Wong, \Layered Coding of Check Images Using Foreground and Background Segmentation," SPIE Conference on Visual Communication and Image Processing, pp. 1040{9, Orlando, FL, March 17{23, A. Netravali and B. Haskell, Digital Pictures, 2nd Ed. Plenum Press, ISO/IEC Draft International Standard 11544, Coded Representation of Picture and Audio Information Progressive Bi-level Image Compression, G. K. Wallace, \The JPEG Still Picture Compression Standard," Communications of the ACM, Vol. 34, pp. 30{ 40, April A. Said and W. A. Pearlman, \A New Fast and Ecient Image Codec Based on Set Partitioning in Hierarchical Trees," IEEE Transactions on Circuits and Systems for Video Technology, Vol. 6, pp. 243{250, June

11 13. ftp: ipl.rpi.edu:pub/ew Code. 14. Portable Video Research Group (PRVG), Stanford University, ftp:havefun.stanford.edu:pub/jpeg/jpegv1.2.1.tar.z. 11

12 List of Tables 1 Bit Allocation Among Dierent Layers for Visually Acceptable Quality (Averaged Over 13 Test Images) Compression Ratios for Dierent Layer Combinations at Visually Acceptable Quality (Averaged Over 13 Test Images) Average Compression Ratios for 13 Test Images by Dierent Algorithms at Visually Acceptable Image Quality List of Figures 1 Original Image of a Check, bits, digitized at 100 dpi Compressed Version of Figure 1 by JPEG, CR = 19.70, PSNR = db Compressed Version of Figure 1 by SPIHT, CR = 20, PSNR = db Original Image of a Check, bits digitized at 200 dpi Compressed Version of Figure 4 by JPEG, CR = 49.6, PSNR = db Compressed Version of Figure 4 by SPIHT, CR = 50.0, PSNR = db Block Diagram of Layered Coding Encoder (a) Illustration of opening with a triangular structuring element; (b) the solid-line is the original signal and the dash-line is the signal after opening operation (a)illustration of closing with a triangular structuring element; (b) the solid-line is the original signal and the dash-line is the signal after closing operation Basic Block Diagram of Background/Foreground Segmentation Procedure to Determine Maximum Width of the Valleys in Each of the Four Directions in a Block Human coding results for Layer II Foreground Segmentation Result for the Image in Figure 1 after the First Pass Foreground Segmentation Result for the Image in Figure 1 after the Second Pass Foreground Segmentation Result for the Image in Figure 4 after the First Pass Foreground Segmentation Result for the Image in Figure 4 after the Second Pass Background of the Image in Figure 1 after Eliminating Foreground Pixels Background of the Image in Figure 1 with Gaps Filled Background of the Image in Figure 4 after Eliminating Foreground Pixels Background of the Image in Figure 4 with Gaps Filled

13 21 Reconstructed Image with Layer I and Layer II Image Compressed by Layered Coding (with JPEG compressed Background), CR = 19.76, PSNR = db Image Compressed by Layered Coding (with SPIHT compressed Background), CR = 19.83, PSNR = db Reconstructed Image with Layer I and Layer II Image Compressed by Layered Coding (with JPEG compressed Background), CR = 50.2, PSNR = db Image Compressed by Layered Coding (with SPIHT compressed Background), CR = 50.1, PSNR = db Image Compressed by JPEG, CR = 9.85, PSNR = db Image Compressed by SPIHT, CR = 8, PSNR = db Image Compressed by JPEG, CR = 28.04, PSNR = db Image Compressed by SPIHT, CR = 28.57, PSNR = db PSNR and Associated Compression Ratios of Dierent Algorithms at 300 dpi PSNR and Associated Compression Ratios of Dierent Algorithms at 200 dpi PSNR and Associated Compression Ratios of Dierent Algorithms at 100 dpi

14 Table 1. Bit Allocation Among Dierent Layers for Visually Acceptable Quality (Averaged Over 13 Test Images). Layer I Layer II Layer III Resolution Byte Percentage Byte Percentage Byte Percentage Total 300 dpi 9, , , , dpi 5, , , , dpi 2, , ,368 Table 2. Compression Ratios for Dierent Layer Combinations at Visually Acceptable Quality (Averaged Over 13 Test Images). Layer I Layer I+II Layer I+II+III Resolution Byte CR Byte CR Byte CR 300 dpi 9, , , dpi 5, , , dpi 2, , , Table 3. Average Compression Ratios for 13 Test Images by Dierent Algorithms at Visually Acceptable Image Quality. Resolution LCj LCw JPEG SPIHT 100 dpi dpi dpi

15 Figure 1. Original Image of a Check, bits, digitized at 100 dpi. Figure 2. Compressed Version of Figure 1 by JPEG, CR = 19.70, PSNR = db. 15

16 Figure 3. Compressed Version of Figure 1 by SPIHT, CR = 20, PSNR = db. Figure 4. Original Image of a Check, bits digitized at 200 dpi. 16

17 Figure 5. Compressed Version of Figure 4 by JPEG, CR = 49.6, PSNR = db. Figure 6. Compressed Version of Figure 4 by SPIHT, CR = 50.0, PSNR = db. 17

18 . Check Image. Foreground Segmentation Foreground Pixel Removing Background Smoothing Binary Foreground Map Background Image Foreground Grayscale Image Foreground Map Coding Gray Level Coding Background Image Coding.. +. Foreground Reconstruction Layer 1 Layer 2 Layer 3 Error Image Coding Layer 4 Figure 7. Block Diagram of Layered Coding Encoder. (a) (b) Figure 8. (a) Illustration of opening with a triangular structuring element; (b) the solid-line is the original signal and the dash-line is the signal after opening operation. 18

19 (a) (b) Figure 9. (a)illustration of closing with a triangular structuring element; (b) the solid-line is the original signal and the dash-line is the signal after closing operation. Input Image Closing + Thresholding Post Processing Foreground Binary Map Figure 10. Basic Block Diagram of Background/Foreground Segmentation. 19

20 maximum width =0; while (not end of block)f counter =0; FoundForeground =0; while (not end of line) f if (FoundForeground == 0) and (current pixel is foreground pixel) f counter =1; Foreground =1; g else f if (FoundForeground ==1) and (gray value is similar to the previous gray value) f counter += 1; g else f if (counter > maximum width)f maximum width = counter; FoundForeground =0; counter =0; g g g advance to next pixel; g if (counter > maximum width) f maximum width = counter; g advance to next line; g Figure 11. Procedure to Determine Maximum Width of the Valleys in Each of the Four Directions in a Block. 20

21 Original Range 15 Range 63 Range Original Range 15 Range 63 Range Original Range 15 Range 63 Range 95 Bytes Bytes 2500 Bytes Check Image # (100 dpi) Check Image # (200 dpi) Check Image # (300 dpi) Figure 12. Human coding results for Layer II. 21

22 Figure 13. Foreground Segmentation Result for the Image in Figure 1 after the First Pass. Figure 14. Foreground Segmentation Result for the Image in Figure 1 after the Second Pass. 22

23 Figure 15. Foreground Segmentation Result for the Image in Figure 4 after the First Pass. Figure 16. Foreground Segmentation Result for the Image in Figure 4 after the Second Pass. 23

24 Figure 17. Background of the Image in Figure 1 after Eliminating Foreground Pixels. Figure 18. Background of the Image in Figure 1 with Gaps Filled. 24

25 Figure 19. Background of the Image in Figure 4 after Eliminating Foreground Pixels. Figure 20. Background of the Image in Figure 4 with Gaps Filled. 25

26 Figure 21. Reconstructed Image with Layer I and Layer II. Figure 22. Image Compressed by Layered Coding (with JPEG compressed Background), CR = 19.76, PSNR = db. 26

27 Figure 23. Image Compressed by Layered Coding (with SPIHT compressed Background), CR = 19.83, PSNR = db. Figure 24. Reconstructed Image with Layer I and Layer II. 27

28 Figure 25. Image Compressed by Layered Coding (with JPEG compressed Background), CR = 50.2, PSNR = db. Figure 26. Image Compressed by Layered Coding (with SPIHT compressed Background), CR = 50.1, PSNR = db. 28

29 Figure 27. Image Compressed by JPEG, CR = 9.85, PSNR = db. Figure 28. Image Compressed by SPIHT, CR = 8, PSNR = db. 29

30 Figure 29. Image Compressed by JPEG, CR = 28.04, PSNR = db. Figure 30. Image Compressed by SPIHT, CR = 28.57, PSNR = db. 30

31 LCj LCw JPEG VTW PSNR Compression Ratio LCj LCw JPEG VTW Check Image # Check Image # Figure 31. PSNR and Associated Compression Ratios of Dierent Algorithms at 300 dpi LCj LCw JPEG VTW LCj LCw JPEG VTW PSNR Compression Ratio Check Image # Check Image # Figure 32. PSNR and Associated Compression Ratios of Dierent Algorithms at 200 dpi. 31

32 LCj LCw JPEG VTW PSNR Compression Ratio LCj LCw JPEG VTW Check Image # Check Image # Figure 33. PSNR and Associated Compression Ratios of Dierent Algorithms at 100 dpi. 32