- An Image Coding Algorithm

Transcription

1 - An Image Coding Algorithm Shufang Wu Friday, June 14,

2 Agenda Overview Discrete Wavelet Transform Zerotree Coding of Wavelet Coefficients Successive-Approximation Quantization (SAQ) Adaptive Arithmetic Coding Relationship to Other Coding Algorithms A Simple Example Experimental Results Conclusion References Q & A 22-2

3 Overview (2-1) Two-fold problem Obtaining best image quality for a given bit rate Accomplishing this task in an embedded fashion What is (EZW)? An embedded coding algorithm 2 properties, 4 features and 2 advantages (next page) What is Embedded Coding? Representing a sequence of binary decisions that distinguish an image from the null image Similar in spirit to binary finite-precision representations of real number 22-3

4 Overview (2-2) - (EZW) 2 Properties Producing a fully embedded bit stream Providing competitive compression performance 4 Features Discrete wavelet transform Zerotree coding of wavelet coefficients Successive-approximation quantization (SAQ) Adaptive arithmetic coding 2 Advantages Precise rate control No training of any kind required 22-4

5 Agenda Overview Discrete Wavelet Transform Zerotree Coding of Wavelet Coefficients Successive-Approximation Quantization (SAQ) Adaptive Arithmetic Coding Relationship to Other Coding Algorithms A Simple Example Experimental Results Conclusion References Q & A 22-5

6 Discrete Wavelet Transform (2-1) Identical to a hierarchical subband system Subbands are logarithmically spaced in frequency Subbands arise from separable application of filters LL2 HL2 LL1 HL1 HL1 LH2 HH2 LH1 HH1 LH1 HH1 First stage Second stage 22-6

7 Discrete Wavelet Transform (2-2) Wavelet decomposition (filters used based on 9-tap symmetric quadrature mirror filters (QMF)) w y Df1 Df1 D 3 2 j f w x In the frequency domain Image wavelet representations 22-7

8 Zerotree Coding (3-1) A typical low-bit rate image coder Large bit budget spent on encoding the significance map Binary decision as to: whether a coefficient has a zero or nonzero quantized value True cost of encoding the actual symbols: Total Cost = Cost of Significance Map + Cost of Nonzero Values 22-8

9 Zerotree Coding (3-2) What is zerotree? A new data structure IF: Parent: To improve the compression of significance maps of A wavelet coefficients The coefficient at a coarse at the scale coarse is scale. Scanning rule: insignificant with a Child: wavelet What is No insignificant? coefficient child respect node x to said a given to be threshold insignificant T, with respect THEN: All coefficients corresponding A is coefficient scanned before to the x same is its parent. spatial location at Zerotree to An a is given based threshold on an empirically T if : true hypothesis All wavelet the element coefficients next finer of a scale zerotree IF of Parent-child dependencies x the of same similar for threshold (descendants < T orientation orientation. T is itself and all of its descendents IF are insignificant the same & ancestors) with It spatial is not location the descendant at finer respect scales of a previously are to T. likely found to be zerotree insignificant root Scanning of the coefficients with for threshold respect to T. T. An element of a zerotree for threshold T A zerotree root Parent-child Scanning dependencies order of the subbands of 22-9

10 Zerotree Coding (3-3) Encoding 22-10

11 SAQ (3-1) Successive-Approximation Quantization (SAQ) Sequentially applies a sequence of thresholds T 0,,T N-1 to determine significance Thresholds Chose so that T i = T i-1 /2 T 0 is chosen so that x j < 2T 0 for all coefficients x j Two separate lists of wavelet coefficients Dominant list Dominant Subordinate list contains: list The Subordinate coordinates list of contains: those coefficients that have not yet been found to be The significant magnitudes in the of those same coefficients relative order that as have the initial been found scan. to be significant

12 SAQ (3-2) Dominant pass During Subordinate a dominant pass pass: coefficients Encoding process with coordinates on the dominant list are compared to During the threshold a subordinate T i to determine pass: their significance, and if significant, all coefficients their on sign. the subordinate list are scanned and the SAQ encoding process: specifications of the magnitudes available to the decoder are FOR I = T refined to an additional bit of 0 TO T precision. N-1 Dominant Pass; Subordinate Pass (generating string of symbols) ; String of symbols is entropy encoded; Sorting (subordinate list in decreasing magnitude); IF (Target stopping condition = TRUE) break; NEXT; 22-12

13 SAQ (3-3) Decoding Each decode symbol, during both passes, refines and reduces the width of the uncertainty interval in which the true value of the coefficient ( or coefficients, in the case of a zerotree root) Reconstruction value Can be anywhere in that uncertainty interval Practically, use the center of the uncertainty interval Good feature Terminating the decoding of an embedded bit stream at a specific point in the bit stream produces exactly the same image that would have resulted had that point been the initial target rate 22-13

14 Adaptive Arithmetic Coding Based on [3], encoder is separate from the model which is basically a histogram During the dominant passes Choose one of four histograms depending on Whether the previous coefficient in the scan is known to be significant Whether the parent is known to be significant During the subordinate passes A single histogram is used 22-14

15 Relationship to Other Coding Algorithms Relationship to Bit Plane Encoding (more general & complex) a) Reduce (all thresholds the width are of the powers largest of uncertainty two and all coefficients interval in all are coefficeints b) Increase integers) the precision further c) Attempt Most-significant to predict insignificance binary digit (MSBD) from low frequency to high Its sign and bit position are measured and encoded during the dominant pass Item PPC EZW 1) 2) 3) Dominant bits First (digits b) to second the left and a) including First the MSBD) a) second b) Measured and No encoded c) during the dominant c) pass Subordinate bits Training (digits to needed the right of the MSBD) No training needed Measured and encoded during the subordinate pass Relationship to Priority-Position Coding (PPC) 22-15

16 Agenda Overview Discrete Wavelet Transform Zerotree Coding of Wavelet Coefficients Successive-Approximation Quantization (SAQ) Adaptive Arithmetic Coding Relationship to Other Coding Algorithms A Simple Example Experimental Results Conclusion References Q & A 22-16

17 A Simple Example (2-1) Only string of symbols shown (No adaptive arithmetic coding) Simple 3-scale wavelet transform of an 8 X 8 image T 0 = 32 (largest coefficient is 63) Example 22-17

18 A Simple Example (2-2) First dominant pass First subordinate pass Example Magnitudes are partitioned into the uncertainty intervals [32, 48) and [48, 64), with symbols 0 and

19 Experimental Results 12-byte header For image Number of Barbara : of wavelet scales Dimensions of the image For number of significant coefficients retained at the same low bit rate: Item Maximum histogram JPEG count for the models EZW in the arithmetic coder Same file Image size mean PSNR & initial lower threshold Looks better Item Other EZW Same Applied PSNR to standard Looks b/w better 8 bpp. test images File size smaller Number retained Less More 512 X 512 Lena image 512 X 512 ( PSNR: Barbara Peak-Signal-to-Noise-Rate image ) (Reason: The zerotree coding provides a much better way of Compared encoding with the JPEG positions of the significant coefficients.) Compared with other wavelet transform coding 22-19

20 Conclusion 2 Properties Producing a fully embedded bit stream Providing competitive compression performance 4 Features Discrete wavelet transform Zerotree coding of wavelet coefficients Successive-approximation quantization (SAQ) Adaptive arithmetic coding 2 Advantages Precise rate control No training of any kind required 22-20

21 References 1. E. H. Adelson, E. Simoncelli, and R. Hingorani, Orthogonal pyramid transforms for image coding, Proc. SPIE, vol.845, Cambridge, MA, Oct. 1987, pp S. Mallat, A theory for multiresolution signal decomposition: The wavelet representation, IEEE Trans. Pattern Anal. Mach. Intell., vol. 11, pp , July I. H. Witten, R. Neal, and J. G. Cleary, Arithmetic coding for data compression, Comm. ACM, vol. 30, pp , June J. Shapiro, Embedded image coding using zerotrees of wavelet coefficients, IEEE Trans. Signal Processing., vol. 41, pp , Dec

22 22-22