JPEG2000 New Standard for Still Image Compression Tao Tao Department of Computer Science University of Central Florida Some of the slides have been adopted from a presentation by Dr T Acharya 1 Outline Introduction Part I JPEG2000 Coding Region of Interest Bit stream ordering Conclusion 2 Introduction 3 Why JPEG2000 5 2 bpp los sy lossless Original image b i t s t r e a m Sub Low er n regio 1 84 bpp res olu tion I RO One component 1 89 bpp 4 Why another still image compression standard To address a number of weakness in the existing JPEG standard To provide a number of new features that available in the JPEG standard Namely Allow efficient lossy and lossless compression within a single unified coding framework Provide superior image quality at low bit rates Support new features such as ROI and a more flexible file format Avoid excessive computational and memory complexity 5 compression standard cont Larger image JPEG does not allow for image greater than 64k by 64k without tiling Transmission in noisy environment JPEG image quality suffers dramatically when biterrors are encountered Computer generated imagery JPEG is optimized for natural imagery only Compound documents JPEG is seldom used in the compression of compound documents because of its poor performance on bi level text imagery 6 7 8 9 10 11 12 13 14 15 16 17 18 JPEG 2000 Standard Part I Core Coding System Part II Extensions Part III Motion JPEG 2000 Part IV Conformance Testing Part V Reference Software Part VI Compound image file format More parts are coming 19 Part I JPEG 2000 20 JPEG 2000 Part I Rate Control Image Multi component transform Discrete Wavelet Quantization Transform Tier 1 Tier 2 Encoder Encoder Coded Image encoder Reconstructed Image Inverse Multi component transform Inverse Wavelet Transform De quantization Tier 1 Tier 2 Decoder Decoder Coded Image decoder 21 JPEG2000 Encoder Subbands tile Subband DWT Subband Image Component code block BPC Subband Code block Tile Context Data Code block BAC Subband Compressed data Bit stream Layer formation and data formatting 22 Multi Component Transform Part I allows color transformation on first three components Reversible Color Transform RCT Irreversible Color Transform ICT Rate Control Image Multi component transform Discrete Wavelet Transform Quantization Tier 1 Encoder Tier 2 Encoder Coded Image Region of Interest 23 Irreversible Color Transform The ICT is nothing more than the classic RGB to YCrCb color space transform 0 587 0 114 R Y 0 299 U 0 16875 0 33126 0 5 G 0 41869 0 08131 B V 0 5 R 1 0 G 1 0 B 1 0 0 0 34413 1 772 1 402 Y 0 71414 U V 0 24 Reversible Color Transform The RCT is simply a reversible integer tointeger approximation to the ICT R 2G B Yr 4 Ur R G Vr B G R Ur G Ur Vr G Yr 4 B Vr G 25 Discrete Wavelet Transform DWT Lifting Scheme Rate Control Image Multi component transform Discrete Wavelet Transform Quantization Tier 1 Encoder Tier 2 Encoder Coded Image Region of Interest 26 Quantization Uniform scalar quantization with deadzone No quantization in lossless mode Quantization rule q sign y y b Rate Control Image Multi component transform Discrete Wavelet Transform Quantization Region of Interest Tier 1 Encoder Tier 2 Encoder Coded Image 27 Deadzone Quantization Small coefficients increase the quality of the image the least and they still mess up the run length encoding as much as big coefficients A deadzone quantizer blocks the small coefficients 28 Entropy Coding Tier 1 coding Bit Plane Coding BPC Tier 2 coding Tag Tree Coding Rate Control Image Multi component transform Discrete Wavelet Transform Quantization Tier 1 Encoder Tier 2 Encoder Coded Image Region of Interest 29 Tier 1 Coding Tier 1 coding is performed on code blocks For each bit plane there are 3 coding passes similar to those in EZW or SPIHT Significance Pass a a r Refinement Pass a a r Clean Up Pass a Example of scan pattern for a 10x5 code block 30 Tier 1 Coding cont Sign Magnitude Representation Samples in a code block Sign 0 0 0 0 1 0 1 0 0 0 0 1 0 0 0 0 4 1 13 15 10 4 0 9 2 5 14 7 11 8 3 15 4 10 2 11 Magnitude 1 4 5 8 13 15 0 9 14 7 15 3 31 Tier 1 Coding cont Significance Each sample in the code block has an associated binary state variable called its significance A sample is significant if it is larger than the current bit plane A sample is predicted to be significant if any of its 8 connected neighbor has been found to be significant Significances are initialized to 0 and may become 1 during the course of the coding of the code block Once the significance for a sample becomes 1 it stays 1 throughout the encoding of the code block 7 binary format 0111 Example 32 Tier 1 Coding cont Significance Pass 33 Tier 1 Coding cont Refinement Pass 34 Tier 1 Coding cont Cleanup Pass 35 Example of Tier 1 Coding 36 Original image 37 2 5 bpp with 4 Level DWT 38 3 5 bpp with 4 Level DWT 39 Compressed at 4 9 bpp with 4 level DWT 40 Region Of Interest 41 DWT ROI mask coefficients contributes only to a specific region A binary mask generated in the wavelet domain for distinction of ROI and Background 42 Original image 43 ROI Example encoded at 5 85 bpp 44 Difference image d 5 45 Bit stream Ordering 46 2 level DWT with 16 Code Blocks 2LL 2HL CB1 CB2 2LH 2HH CB3 CB4 CB9 CB10 1LH CB11 CB5 CB6 1HL CB7 CB8 CB13 CB14 1HH CB12 CB15 CB16 47 Assume maximum 4 Bit Planes for each Code Block BP 4 C C C C C C C C C C C C C C C C BP 3 S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C BP 2 S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C BP 1 S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C S M C CB1 CB2 CB3 CB4 CB5 CB6 CB7 CB8 CB9 CB19 CB11 CB12 CB13 CB14 CB15 CB16 2LL 2HL 2LH 2HH 1HL 1LH 1HH S Significance propagation pass M Magnitude refinement pass C Clean up pass 48 Bit stream with lower resolution image BP 4 C C …