Module 6 STILL IMAE COMPRESSION STANDARDS
Lesson 17 JPE-2000 Achitectue and Featues
Instuctional Objectives At the end of this lesson, the students should be able to: 1. State the shotcomings of JPE standad. 2. State the scope and objectives of JPE-2000. 3. Name the main application aeas of JPE-2000. 4. List the equiements of its majo applications. 5. Explain the majo building blocks of JPE-2000. 6. Define tiling and its significance. 7. Distinguish between evesible and ievesible wavelet tansfoms. 8. Pefom peiodic extension of signals. 9. Distinguish between convolution based filteing and lifting based filteing. 10. Show the elations fo evesible and ievesible component tansfomations. 11. Define pecincts and packets fo JPE-2000 bit steam. 17.0 Intoduction In lesson-16, we had discussed in details about the JPE standad. This was in fact the fist eve intenational standad on continuous tone still image compession and coding. Despite its vey impessive pefomance and the massive populaity it gained in stoage of still images and tansmission though intenet, some shotcomings wee felt as the technology futhe matued. Fist, JPE encoded images show sevee blocking atifacts at vey low bit ate. This poblem basically oiginates fom the use of block-based DCTs, which we had discussed ealie. Moeove, JPE povides insufficient suppot fo spatial and SNR (quality) scalability. JPE lacks object-based and egion-based epesentation and has no suppot fo eo esiliency. Subsequent to the acceptance of JPE as a standad, the technology petaining to image compession and coding matued futhe, especially in the field of wavelet tansfoms and subband coding, which we had discussed in lessons-10 to 15. The shotcomings of JPE, matuity of wavelet tansfoms and subband coding and polifeation of new application aeas pompted the intenational community to wok on a moe advanced still image coding standad, known as JPE-2000, whose wok was fist initiated in 1996.
In this lesson, we ae going to pesent the majo application aeas of JPE-2000 and thei equiements, the achitectue of the coding engine and the encoding concepts. JPE-2000 encoding is based on EBCOT wavelet coding technique, which we have aleady discussed in lesson-15. Hence, the focus of this lesson will not be on the coding algoithm, but moe on its achitectual aspects. Because of seveal added featues not found in ealie coding standads, impoved capabilities and vast futue potentials, we shall devote two moe lessons, viz. lesson-18 and lesson-19, in addition to the pesent one to discuss vaious aspects of JPE-2000 coding. 17.1 Scope and objectives of JPE-2000 The JPE-2000 standads wee intended fo use in diffeent types of still images, such as bi-level, gay-scale, colo (o multi-component) with diffeent chaacteistics, such as natual images, scientific, medical, emote sensing, text, endeed gaphics etc and on diffeent imaging models, such as client/seve, eal-time tansmission, image libay achival, limited buffe and bandwidth esouces etc all within a unified famewok. The objective was to make it 30% moe efficient than the baseline JPE. The standadization committee developed the Pat-I of JPE-2000 standad to make it oyalty fee, so as to pomote its wide usage. 17.2 Applications of JPE-2000 and thei equiements The JPE-2000 standads have seveal applications and moe ae emeging with time. Some of the majo application aeas ae: Intenet Colo facsimile Pinting Scanning Digital photogaphy Remote Sensing Mobile Medical imagey Digital libaies and achives E-commece Each application aea has some equiements which the standad should fulfill. The equiements fom diffeent application aeas may be summaized as follows:
Impoved low bit-ate pefomance: It should give acceptable quality below 0.25 bpp. Netwoked image delivey and emote sensing applications have this equiements. Lossless and lossy compession: Lossless compession should be included in pogessive decoding. It should suppot embedded bit-steam that allow pogessive lossy to lossless build-up. Pogessive tansmission: The standad should allow pogessive tansmission that allows images to be econstucted with inceasing pixel accuacy and esolution. Region of Inteest Coding: It should pefeentially allocate moe bits to the egions of inteest (ROIs) as compaed to the non-roi ones. Random code steam access: The use defined ROIs in an image should be andomly accessible. Eo esilience: The standad should povide eo esiliency, especially fo wieless communication channels. Open achitectue: It should suppot open achitectue to optimize the system fo diffeent image types and applications. The decode should have only the coe tool set and a pase that undestands the coe steam. If necessay, the unknown tools could be equested by the decode. Content based desciption: Finding the desied image fom a lage achive of images is a challenging task. This has applications in medical images, foensic, digital libaies etc. These issues ae being addessed by MPE-7. Image Secuity: This is anothe majo equiement to potect the intellectual popety ights of the images. Digital images can be potected using watemaking, labeling, stamping, encyption etc. Continuous-tone and bi-level image compession: Peviously, we had diffeent standads fo bi-level images (JBI standads) and continuoustome images (JPE standad). A equiement was felt that a single standad should addess both these domains, including the colo images and encompass applications involving compound images with ovelaying texts. Sequential one-pass decoding: The decoding should be done in a single sequential pass and suppot inteleaved and non-inteleaved mode.
Side-channel suppot: The standad should suppot alpha planes and tanspaency planes. All these equiements, oiginating fom diffeent application domains wee addessed and incopoated in the JPE-2000 standad. 17.3 Achitectue of JPE-2000 Fig. 17.1: Block Diagam of JPE-2000 encode. Fig.17.1 shows the block-diagam of JPE-2000 encode and fig.17.2 shows its coesponding decode. Fig. 17.2: Block Diagam of JPE-2000 decode The block diagams ae not significantly diffeent fom those of JPE, except fo the fact that the tansfom used is Discete Wavelet Tansfom (DWT), athe than the DCT. Befoe applying the DWT, the souce image is divided into components (colos) and each component is divided into tiles (to be descibed in Section-17.4) which ae compessed independently, as though they ae independent images and all the samples in the tile ae DC level shifted. The DWT coefficients in diffeent subbands ae quantized and then composed into an embedded bit-steam following the EBCOT algoithm, pesented in lesson-15. The embedded bit-steam is composed of quality layes so as to offe both esolution and SNR scalability. The entopy coding is based on a contextadaptive aithmetic code, known as MQ code, which has eo coection capability. The encoding pocess may be summaized as follows:
The souce image is decomposed into components, such as RB o YUV. The image and its components ae divided into ectangula tiles, which fom the basic unit to be encoded. DWT at diffeent esolution levels is applied on each tile to compose the subbands. The DWT may be evesible o ievesible, to be descibed in Section-17.5. The subband coefficients ae quantized and collected into ectangula aays of code blocks (descibed in lesson-15, Section-15.3). The bit-planes of the code-blocks ae aithmetic coded in multiple passes with factional bit-plane concept of EBCOT algoithm (efe lesson-15). The encoding is done in such a way that the Regions of Inteest (ROI) get highe quality, as compaed to the othe aeas (to be discussed in lesson- 18). Makes ae intoduced in the bit-steam fo eo esilience (to be discussed in lesson-19). 17.4 Tiling and its significance Befoe applying the DWT, the image and its components ae divided into smalle non-ovelapping blocks, known as tiles, which can be coded independently, as if each tile is an independent image. All opeations, such as component mixing, DWT, quantization and entopy coding ae theefoe done independently fo each tile. Tiling has the advantage of educing memoy equiements fo DWT and its pocessing and is amenable to paallelization. Moeove, tiles may be independently accessed and used fo decoding specific pats of the image, athe than the complete one. The tile may be as lage as the entie image size (that is, single tile) o of smalle patitions, such as 256x256, 128x128 etc. In tems of PSNR, tiling degades the pefomance, as compaed to no tiling and smalle tile sizes lead to tiling atifacts. 17.5 Wavelet Filtes The JPE-2000 standad suppots lossy as well as lossless encoding. Two types of wavelet filtes ae included in Pat-I of the standad
(a) Ievesible, whee exact econstuction will not be possible at the decode and is used fo lossy encoding. This is implemented using a 9/7 Daubechies filte, whose analysis and synthesis filte coefficients ae shown in Table-17.1. (b) Revesible, whee exact econstuction at the decode is possible and is theefoe included fo lossless JPE-2000. This is implemented using a 5/3 filte, whose analysis and synthesis filte coefficients ae shown in Table-17.2. Analysis Filte Coefficients i Low-pass filte h l ( i) High-pass filte h u ( i) 0 0.6029490182363579 1.115087052456994 ± 1 0.2668641184428723-0.5912717631142470 ± 2-0.07822326652898785-0.05754352622849957 ± 3-0.01686411844287495 0.09127176311424948 ± 4 0.02674875741080976 i Synthesis Filte Coefficients i i Low-pass filte h l ( ) High-pass filte ( i) 0 1.115087052456994 0.6029490182363579 ± 1 0.5912717631142470-0.2668641184428723 ± 2-0.05754352622849957-0.07822326652898785 ± 3-0.09127176311424948 0.01686411844287495 ± 4 0.02674875741080976 Table-17.1 Daubechies 9/7 analysis and synthesis filte coefficients Analysis Filte Coefficients Low-pass filte i h l () High-pass filte i h u ( ) g l ( ) h u Synthesis Filte Coefficients Low-pass filte High-pass filte i i 0 6/8 1 1 6/8 ± 1 2/8-1/2 1/2-2/8 ± 2-1/8-1/8 g u () Table-17.2 5/3 Analysis and synthesis filte coefficients
17.5.1 Peiodic Extension of Signals: Two filteing appoaches, namely convolution and lifting based (to be discussed in Section-17.5.1) filtes ae used in JPE-2000 standad. Both these appoaches fail at the signal boundaies, since coesponding signal samples ae not available fo all the filte coefficients. One can conside the non-existent samples as zeos, but this leads to unsatisfactoy esults because of signal discontinuities. A bette appoach is to extend the signal at both the boundaies by including the mio images of the samples about the boundaies. Fig. 17.3: Peiodic symmetic extension of 1-d signal S(0),S(1) S(8). Fig. 17.3 illustates the signal extension fo a one-dimensional example of nine samples: s( 0), s( 1 ), s( 2),, s( 8), whee s( 0) and s ( 8) seve as bounday samples. s 0 and to the ight of s 8 as The samples ae extended to the left of ( ) s s () i = s( i), i = 1, 2, ( 8 + i) = s( 8 i) i = 1,2,3, The extent to which the samples should be extended depends on the filte length. 17.5.2 Convolution and Lifting Based filtes: The convolution based filteing pefoms a dot poduct between the two filte masks (fo low-pass and high-pass filtes) and the extended 1-D signal. In lifting based filteing, the samples ae split into odd and even goups. The odd samples ae updated with a weighted sum of even sample values and the even samples ae updated with a weighted sum of odd sample values. Fo the evesible filtes (lossless case), the esults ae ounded to intege values. The lifting based filteing fo the 5/3 analysis filte is achieved as follows: y y ( 2n + 1) = x ( 2n + 1) ( 2n) = x ( 2n) ext ext + ( 2n) + x ( 2n 2) xext ext + 2 ( 2n 1) + y( 2n + 1) y + 4 2 ()
whee is the extended input signal and y is the output signal. x ext 17.6 Revesible and Ievesible Component Tansfomations JPE-2000 suppots two diffeent component tansfomations fo multicomponent images: Ievesible Component Tansfomation (ICT) used fo lossy coding only. It may only be used with the 9/7 ievesible wavelet tansfom. Revesible Component Tansfomation (RCT) used fo lossless as well as fo lossy coding. It may only be used with the 5/3 evesible wavelet tansfom. The fowad (RB to YUV) and the invese ICT (YUV to RB) tansfomations ae given by = B R V U Y 0.08131 0.41869 0.5 0.5 0.33126 0.16875 0.114 0.587 0.299 = V U Y B R 0 1.772 1.0 0.71414 0.34413 1.0 1.402 0 1.0 The fowad and the invese RCT tansfomations ae given by + + = B R B R V U Y 4 2 + + + = V U V U Y B R 4
The pefomance compaisons between lossless compession (RCT + 5/3 evesible filte) and lossy compession (ICT + 9/7 ievesible filte) at compaable bit ate indicates significantly bette pefomance fo the latte, as compaed to the fome. It may be noted that the sub-sampling of the chominance components, as shown in Section-16.4 of lesson-16 is not ecommended fo JPE-2000, since a 2:1 sub-sampling in hoizontal and vetical diections may be obtained by simply discading the HL, LH and HH subbands of a component s wavelet decomposition and etaining all othe subbands. 17.7 Pecincts and packets fo JPE-2000 bit-steam The aithmetically coded quantized wavelet coefficients ae aanged in packets and the packet patition locations ae efeed to as pecincts. Afte quantization, each subband is divided into some non-ovelapping ectangles. Thee spatially consistent ectangles - one fom each subband at each esolution level compise a packet patition location o pecinct, as illustated in fig.17.4. Each pecinct is futhe divided into code-blocks (discussed in Section-15.3 of lesson-15), which fom the input to the aithmetic code. The data epesenting a specific tile, laye (efe to lesson-15), component, esolution and pecinct appeas in the code steam in a contiguous segment, called a packet. The data in a packet is odeed accoding to the subbands LL, HL, LH and HH and within each subband, the codeblocks ae aanged in aste scanning ode, confined within the bounds of the coesponding pecinct.