Pioneering VDT Image Compression using Block Coding
0 likes43 views
This paper proposes a block-based compression system for VDT images that divides images into blocks that are classified into four types (smooth, text, hybrid, picture) based on histogram and gradient properties. Different coding algorithms are used for each block type to efficiently compress based on their characteristics. Experimental results show the proposed system outperforms JPEG and DjVu in terms of quality and compression performance for VDT images, with comparable complexity to JPEG.
![IEE Proceedings - Vision, Image and Signal Processing, 152 (4), August 2005, pp. 513-518
513
Pioneering VDT Image Compression using Block Coding
P.S.Jagadeesh Kumar
Associate Professor, Department of Compute Science
University of Cambridge, United Kingdom
Abstract: This paper patronage a unique block-built compression system for VDT images. The
images are alienated to blocks, which are pigeon-holed into four dissimilar type blocks namely
smooth blocks, text blocks, hybrid blocks and picture blocks with a dynamic block-built taxonomy
system. Four dissimilar coding systems are castoff for respective block type bestowing to their
disparate arithmetical chattels; to exploit their compression recital. Imitations demonstrate that the
suggested exertion has subordinate intricacy than DjVu with substantially enhanced chromatic
eminence at high bit rate, and it also outclasses the standard lossy image coding scheme JPEG.
Keywords: Block-built Compression, Block Taxonomy, Coding algorithms, VDT images.
1 Introduction
With the prevalent of digital diplomacies such as digital cameras, private PCs, ever
more VDT images, comprehending text, graphics and natural images, are prevailing in
digital systems such as screen images, web pages. The compassion of human eyes for
natural image and text is not the same. The quality prerequisite of VDT image coding
[1] is dissimilar from overall image coding since users cannot assent the quality if text is
not vibrant enough to identify. How to compress VDT image is a stiff delinquent. A
plenty of algorithms have been premeditated to compress images with dissimilar types.
The Lempel-Ziv algorithm [2] is deliberate in compressing pure text images, which only
have text on the pure color background in the entire image. The JPEG algorithm [3] is
decorous for pure picture images which do not have any text in the entire images, but
has debauched enactment on pure text images. Quite a few algorithms were also
suggested to compress VDT images. One type of maneuvers for VDT images is layered
coding. Utmost layered coding algorithms use the typical three layer mixed raster
content exemplification. One prevalent method is DjVu, which practices a wavelet-
based codec (IW44) for background and foreground, and JB2 for mask layer.
Conversely, the intricacy of mask generating and IW44 wavelet transform is great,
which makes DjVu not fit for real time application. Moreover, DjVu shows debauched
concert on pure text images [4,5,6].
Block-built methodologies for VDT images are also willful for their low intricacy.
Said et al. anticipated a simple blocked-based scheme, which compresses text blocks
using JPEG-LS, picture blocks using JPEG [7]. Conversely, it nosedives to knob the
hybrid blocks, which comprises mixed text and pictures. In text area, there are sturdy
edges which cannot finger commendably by DCT centered coding such as JPEG. In this
paper, extant an unusual system, which can progressively compress images with diverse
content types, such as pure text images, pure picture images, and VDT images. Four
coding algorithms are premeditated for those blocks with disparate types based on a
dissolute block taxonomy method.
2 Block Coding System
2.1 Block-built compression
The VDT image is foremost divided into 16x16 blocks. The 16x16 block
dissection is done in-order to upturn the exactitude of segmentation as well as to
expedite superior taxonomy of blocks. Further 32x32 blocks may upsurge the intricacy
of block taxonomy. At that time, blocks are categorized into four types: smooth, text,
hybrid and picture according to their different arithmetic physiognomies. Blocks of
diverse type will be compressed with altered algorithms. The block type map is
compressed using an arithmetic coder.](https://image.slidesharecdn.com/pioneeringvdtimagecompressionusingblockcoding-210208064415/85/Pioneering-VDT-Image-Compression-using-Block-Coding-1-320.jpg)
![IEE Proceedings - Vision, Image and Signal Processing, 152 (4), August 2005, pp. 513-518
515
Fig. 2. Color Quantization Algorithm
Each pixel’s color in the text block is initially transformed to color index. The chief
colors are indexed by 0, 1, 2 and 3 correspondingly. All the other colors are rehabilitated
to index 4. The chief colors in each block are chronicled.
Fig. 3. Text block coding contexts
The text block index is skimmed and compressed in a raster scanning order and the
present pixel index is coded based on its causal neighbors as shown in Fig.3 to endeavor the
spatial significance in order to advance coding efficacy. “X” is present pixel to be implied.
Each neighbor may be five different index values; there are total 5 125 3 contexts for coding
codin the present pixel X. The present pixel index is coded by an arithmetic coder using the
context specified by its three casual neighbors [NW, N, and W]. If the present pixel index is
4, the pixel value is also coded using an arithmetic coder.
Fig. 4. Text Block Coding
2.5 Hybrid Block Coding
Hybrid blocks comprise variegated text and pictures. There are robust high frequency
signals owing to text edges. DCT transform is only operative to condense the vigor of
low frequency signals, so the vigor in DCT domain of hybrid block is very dissimilar
and inflexible to code. When Hybrid blocks are compressed with DCT block transform
based coding such as JPEG, it will agonize from ringing effects about the text due to
huge quantization step for those high frequency constituents. Wavelet-based systems
such as JPEG-2K nosedive to compress hybrid blocks effectually.](https://image.slidesharecdn.com/pioneeringvdtimagecompressionusingblockcoding-210208064415/85/Pioneering-VDT-Image-Compression-using-Block-Coding-3-320.jpg)
![IEE Proceedings - Vision, Image and Signal Processing, 152 (4), August 2005, pp. 513-518
518
Fig. 9. Visual quality comparisons of portion of reconstructed images:
Top (VDT2) and Bottom (VDT3).
The proposed system is of low complexity since the algorithms castoff has less time
complexity. The taxonomy algorithm just customs the histogram of a block; the hybrid
coder practices haar wavelet; the text coder practices context-based arithmetic coding;
the most reckoning lavish part of scheme is picture coding, which use a JPEG-like
algorithm. In the vilest, the proposed system has analogous intricacy with JPEG.
4 Conclusion
In this exertion, a Block-built compression system for VDT image coding is acquiesced.
A profligate block taxonomy algorithm is suggested and four compression algorithms are
prudently deliberated for blocks of diverse kinds. The proposed system accomplishes
analogous or improved coding recital, on pure text images and pure picture images, to LZW
and JPEG respectively. The system outclasses DjVu on VDT images at high bitrate.
References
[1] “Coding of still images”, ITU-T, ISO/IEC SG 8 FCD 15444-6, 2004.
[2] J.Ziv and A. Lempel, “A universal algorithm for data compression”, IEEE
Transaction on Information Theory, IT-23(3), pp.337-343, May 2003.
[3] W. P. Pennebaker, J. L. Mitchell, “JPEG: Still image compression standard”, Van
Nostrand Reinhold, 2004.
[4] “Mixed Raster Content (MRC)”, ITU-T Recommendation T.44, Study Group-8
Contributions, 2001.
[5] P. Haffner, L. Bottou, P.G. Howard, P. Simard, Y. Bengio, Y. LeCun, “High Quality
document image compression with DjVu”, Journal of Electronic Imaging, pp.410-
425, July, 1998.
[6] B.-f Wu, C.-C Chiu and Y. –L Chen, “Algorithms for compressing compound
document images with large text/background overlap”, IEEE Proceedings on
visual image signal process, Vol. 151 No. 6, pp.453-459 December 1997.
[7] A.Said and A.Drukarev, “Simplified segmentation for compound image
compression”, Proceeding of ICIP’ 2004, pp.229-233.
BIOGRAPHY
P.S.Jagadesh Kumar is working as Associate Professor in the Department of
Computer Science, University of Cambridge, United Kingdom. He received his B.E
degree from University of Madras in EEE discipline by the year 1999. He obtained
his M.E degree in 2004 with specialization in CSE from Annamalai University,
Chidambaram, India.](https://image.slidesharecdn.com/pioneeringvdtimagecompressionusingblockcoding-210208064415/85/Pioneering-VDT-Image-Compression-using-Block-Coding-6-320.jpg)
Pioneering VDT Image Compression using Block Coding
- 1. IEE Proceedings - Vision, Image and Signal Processing, 152 (4), August 2005, pp. 513-518 513 Pioneering VDT Image Compression using Block Coding P.S.Jagadeesh Kumar Associate Professor, Department of Compute Science University of Cambridge, United Kingdom Abstract: This paper patronage a unique block-built compression system for VDT images. The images are alienated to blocks, which are pigeon-holed into four dissimilar type blocks namely smooth blocks, text blocks, hybrid blocks and picture blocks with a dynamic block-built taxonomy system. Four dissimilar coding systems are castoff for respective block type bestowing to their disparate arithmetical chattels; to exploit their compression recital. Imitations demonstrate that the suggested exertion has subordinate intricacy than DjVu with substantially enhanced chromatic eminence at high bit rate, and it also outclasses the standard lossy image coding scheme JPEG. Keywords: Block-built Compression, Block Taxonomy, Coding algorithms, VDT images. 1 Introduction With the prevalent of digital diplomacies such as digital cameras, private PCs, ever more VDT images, comprehending text, graphics and natural images, are prevailing in digital systems such as screen images, web pages. The compassion of human eyes for natural image and text is not the same. The quality prerequisite of VDT image coding [1] is dissimilar from overall image coding since users cannot assent the quality if text is not vibrant enough to identify. How to compress VDT image is a stiff delinquent. A plenty of algorithms have been premeditated to compress images with dissimilar types. The Lempel-Ziv algorithm [2] is deliberate in compressing pure text images, which only have text on the pure color background in the entire image. The JPEG algorithm [3] is decorous for pure picture images which do not have any text in the entire images, but has debauched enactment on pure text images. Quite a few algorithms were also suggested to compress VDT images. One type of maneuvers for VDT images is layered coding. Utmost layered coding algorithms use the typical three layer mixed raster content exemplification. One prevalent method is DjVu, which practices a wavelet- based codec (IW44) for background and foreground, and JB2 for mask layer. Conversely, the intricacy of mask generating and IW44 wavelet transform is great, which makes DjVu not fit for real time application. Moreover, DjVu shows debauched concert on pure text images [4,5,6]. Block-built methodologies for VDT images are also willful for their low intricacy. Said et al. anticipated a simple blocked-based scheme, which compresses text blocks using JPEG-LS, picture blocks using JPEG [7]. Conversely, it nosedives to knob the hybrid blocks, which comprises mixed text and pictures. In text area, there are sturdy edges which cannot finger commendably by DCT centered coding such as JPEG. In this paper, extant an unusual system, which can progressively compress images with diverse content types, such as pure text images, pure picture images, and VDT images. Four coding algorithms are premeditated for those blocks with disparate types based on a dissolute block taxonomy method. 2 Block Coding System 2.1 Block-built compression The VDT image is foremost divided into 16x16 blocks. The 16x16 block dissection is done in-order to upturn the exactitude of segmentation as well as to expedite superior taxonomy of blocks. Further 32x32 blocks may upsurge the intricacy of block taxonomy. At that time, blocks are categorized into four types: smooth, text, hybrid and picture according to their different arithmetic physiognomies. Blocks of diverse type will be compressed with altered algorithms. The block type map is compressed using an arithmetic coder.
- 2. IEE Proceedings - Vision, Image and Signal Processing, 152 (4), August 2005, pp. 513-518 514 2.2 Block taxonomy Algorithm The block taxonomy algorithm is habitually a reckoning arduous job. Here proposed a nippy and operative taxonomy algorithm based on two skins: histogram and ascent of the block. The pixels of per capita block are primarily congregated into three classes: low-ascent pixels, mid-ascent pixels and high-ascent pixels according to pixel’s ascent value. Then the histogram dispersal of per capita pixel group is figured. The emblematic ascent-histogram dissemination is shown in Fig.1. As witnessed in Fig.1, blocks of dissimilar type spectacles unlike ascent-histogram distributions. At that moment categorized the blocks into four types: smooth, text, hybrid and picture blocks based on ascent-histogram dissemination. The smooth blocks archetypally comprehend low-ascent pixels and show one peak at the low-ascent histogram. The text blocks constantly show several peaks at the low-ascent and high-ascent histograms. Merely a little mid-ascent pixels can be observed in text blocks. If the block comprehends large numbers of high-ascent pixels and mid-ascent pixels, it will be recognized as hybrid block. The blocks mostly consisting of mid-ascent pixels are blatant as picture blocks. Here T1~T7 are thresholds for arbitrating the block types. The low, mid and high ascent pixel values as well as threshold values are kept undisclosed for intellect reasons due to their inconsistent nature and are prevailed in range for ease. Fig. 1. The three histograms of text block (left), picture block (mid) and hybrid block (right) The blocks of dissimilar type are discrete in general and have dissimilar information disseminations. Smooth blocks are plane and subjugated by one benevolent of color. Text blocks are denser in spatial domain than in DCT domain. The influence of picture blocks is mainly strenuous on low frequency coefficients when they are DCT transmuted. Hybrid blocks, covering mixed text and picture images, cannot be efficiently epitomized both in spatial and frequency domain. Four coding algorithms are prudently intended to compress blocks of dissimilar types effectually. 2.3 Smooth Block Coding The coding of smooth blocks is candid. Smooth blocks are controlled by one color and their gray level range is restricted to the assumed threshold. Every colors in smooth blocks are quantized to the utmost recurrent color, which is veiled by an arithmetic coder. 2.4 Text Block Coding The text blocks are archetypally subjugated by numerous chief colors. The colors with incidence above the assumed threshold are preferred as chief colors. If close by over and above four colors sustaining beyond obligation, solitary the first four colors with prime number in luminance histogram will be preferred as chief colors. The colors adjacent to chief colors inside the assumed threshold are quantized to their conforming chief colors. The color quantization algorithm can be described as shown in Fig.2.
- 3. IEE Proceedings - Vision, Image and Signal Processing, 152 (4), August 2005, pp. 513-518 515 Fig. 2. Color Quantization Algorithm Each pixel’s color in the text block is initially transformed to color index. The chief colors are indexed by 0, 1, 2 and 3 correspondingly. All the other colors are rehabilitated to index 4. The chief colors in each block are chronicled. Fig. 3. Text block coding contexts The text block index is skimmed and compressed in a raster scanning order and the present pixel index is coded based on its causal neighbors as shown in Fig.3 to endeavor the spatial significance in order to advance coding efficacy. “X” is present pixel to be implied. Each neighbor may be five different index values; there are total 5 125 3 contexts for coding codin the present pixel X. The present pixel index is coded by an arithmetic coder using the context specified by its three casual neighbors [NW, N, and W]. If the present pixel index is 4, the pixel value is also coded using an arithmetic coder. Fig. 4. Text Block Coding 2.5 Hybrid Block Coding Hybrid blocks comprise variegated text and pictures. There are robust high frequency signals owing to text edges. DCT transform is only operative to condense the vigor of low frequency signals, so the vigor in DCT domain of hybrid block is very dissimilar and inflexible to code. When Hybrid blocks are compressed with DCT block transform based coding such as JPEG, it will agonize from ringing effects about the text due to huge quantization step for those high frequency constituents. Wavelet-based systems such as JPEG-2K nosedive to compress hybrid blocks effectually.
- 4. IEE Proceedings - Vision, Image and Signal Processing, 152 (4), August 2005, pp. 513-518 516 Whereas hybrid blocks are compressed with document image algorithms, the coding recital is too squat to be satisfactory. One elucidation to this delinquent is layered coding such as DjVu. The text and pictures are detached into dissimilar layers and self-reliantly coded. Proposed a haar wavelet based coding algorithm for hybrid blocks. As we know, short wavelet bases are obliging to moderate the ringing effect around text (edge), and longer bases are virtuous to advance the coding recital of the picture images. As a quid pro quo between two chucks, haar wavelet is designated to code hybrid blocks. Though haar wavelet’s poor recital on pure picture images, it can meritoriously confiscate the ringing effect on text images. Its coding recital outclasses other coding algorithms both in PSNR and visual quality. The hybrid block is first malformed with haar wavelet. Now that only one level haar wavelet transform is castoff, subsequently multilevel haar wavelet transform will produce long wavelet bases not appropriate for hybrid blocks. The wavelet coefficients are then coded by a simple arithmetic coder. The coefficients of dissimilar sub bands are veiled using unlike contexts. The simple haar wavelet algorithm can suggestively advance the visual quality and PSNR of images with hybrid blocks. The visual quality of rebuilt images of three methods is equated in Fig.5: coded by one level haar wavelet, three levels haar wavelet and DCT. The PSNR-bitrate curves of the three methods are shown in Fig.6. It is palpable that one level haar wavelet attains the finest coding recital for hybrids blocks. Fig. 5. Visual quality comparisons of reconstructed images Fig. 6. PSNR-bitrate curves for hybrid blocks of three methods: haar one level, haar three level and DCT. 2.6 Picture Block Coding JPEG has been evidenced to be an effective low complexity algorithm to compress picture images. Here proposed a JPEG-like algorithm to compress the picture blocks. The dissimilarity is that it is needed to gambol the blocks of other types. As expected, the system achieves analogous coding recital to jpeg on pure photographic images with trivial overhead of blocks type map.
- 5. IEE Proceedings - Vision, Image and Signal Processing, 152 (4), August 2005, pp. 513-518 517 3 Experiment results Five images are verified in our tryouts: VDT1, VDT2, VDT3, picture1 and text1, which are revealed in Fig.7. VDT1~VDT3 are instances of typical VDT images with mixed text and pictures; Text1 is a document image; Picture1 is the sample of pure picture images. The coding efficacy between the proposed system and JPEG are equated in Fig.8 for three of the VDT images. The proposed system undoubtedly outclasses jpeg for all the images. In the pre-eminent circumstance, the proposed system outperforms JPEG up to 7db for VDT2, which comprises a ration of hybrid blocks. Fig.9 relates the visual quality of rebuilt images coded by JPEG, DjVu, and the proposed system at the similar bitrate. It is apparent that the proposed system accomplishes much improved visual quality than JPEG and also outclasses DjVu. The ringing affects round text in text blocks and hybrid blocks are efficaciously detached by the proposed system. Fig. 7. Test images: (a) VDT1, (b) VDT2, (c) VDT3, (d) Picuture1 and (e) Text1 Fig. 8. PSNR-bitrate curves for VDT1, VDT2 and VDT3 of BB vs. JPEG
- 6. IEE Proceedings - Vision, Image and Signal Processing, 152 (4), August 2005, pp. 513-518 518 Fig. 9. Visual quality comparisons of portion of reconstructed images: Top (VDT2) and Bottom (VDT3). The proposed system is of low complexity since the algorithms castoff has less time complexity. The taxonomy algorithm just customs the histogram of a block; the hybrid coder practices haar wavelet; the text coder practices context-based arithmetic coding; the most reckoning lavish part of scheme is picture coding, which use a JPEG-like algorithm. In the vilest, the proposed system has analogous intricacy with JPEG. 4 Conclusion In this exertion, a Block-built compression system for VDT image coding is acquiesced. A profligate block taxonomy algorithm is suggested and four compression algorithms are prudently deliberated for blocks of diverse kinds. The proposed system accomplishes analogous or improved coding recital, on pure text images and pure picture images, to LZW and JPEG respectively. The system outclasses DjVu on VDT images at high bitrate. References [1] “Coding of still images”, ITU-T, ISO/IEC SG 8 FCD 15444-6, 2004. [2] J.Ziv and A. Lempel, “A universal algorithm for data compression”, IEEE Transaction on Information Theory, IT-23(3), pp.337-343, May 2003. [3] W. P. Pennebaker, J. L. Mitchell, “JPEG: Still image compression standard”, Van Nostrand Reinhold, 2004. [4] “Mixed Raster Content (MRC)”, ITU-T Recommendation T.44, Study Group-8 Contributions, 2001. [5] P. Haffner, L. Bottou, P.G. Howard, P. Simard, Y. Bengio, Y. LeCun, “High Quality document image compression with DjVu”, Journal of Electronic Imaging, pp.410- 425, July, 1998. [6] B.-f Wu, C.-C Chiu and Y. –L Chen, “Algorithms for compressing compound document images with large text/background overlap”, IEEE Proceedings on visual image signal process, Vol. 151 No. 6, pp.453-459 December 1997. [7] A.Said and A.Drukarev, “Simplified segmentation for compound image compression”, Proceeding of ICIP’ 2004, pp.229-233. BIOGRAPHY P.S.Jagadesh Kumar is working as Associate Professor in the Department of Computer Science, University of Cambridge, United Kingdom. He received his B.E degree from University of Madras in EEE discipline by the year 1999. He obtained his M.E degree in 2004 with specialization in CSE from Annamalai University, Chidambaram, India.