A Gradient Based Predictive Coding for Lossless Image

A Gradient Based Predictive Coding for Lossless Image Compression Source: IEICE Transactions on Information and Systems, Vol. E 89 -D, No. 7, July 2006. Authors: Haijiang Tang and Sei-ichiro Kamata Speaker: Chia-Chun Wu Date: 2006/10/19 1

Outline o o o o 1. 2. 3. 4. 5. 6. 7. Lossless image compression Predictive coding LOCO-I (JPEG-LS) CALIC The proposed scheme Experimental results Conclusions 2

1. Lossless image compression o Lossless: reconstruct the coded image identically to the original image o Applications: • • • Medical imaging Remote sensing Fax Image archiving Art work preserving … 3

2. Predictive coding o Practice: The value of a pixel can be accurately predicted using a simple predictor of previously observed neighbor pixels. c b a x 4

LOCO-I: Low complexity lossless compression for images 3. LOCO-I (JPEG-LS) o median edge detector Example: 60 105 100 105 60 50 102 60 50 Original image 105 100 105 Predictive values e = {+5, +2, 45} 5

CALIC: Context-based, adaptive, lossless image coder 4. CALIC o gradient adjusted predictor d g h c b i a x Causal template 6

4. CALIC (cont. ) o gradient adjusted predictor d g h c b i a x Causal template Sharp horizontal Example: 40 30 15 45 20 20 25 102 105 100 dv-dh=105 -8=97>80 e = -5 Horizontal Weak horizontal 40 55 50 45 50 65 54 102 105 100 dv-dh=69 -29=40 >32 =(86+105)/2=96 e = +4 55 60 60 60 100 50 45 50 55 100 dv-dh=70 -60=10 >8 =(3*39+55)/4 = 43 e = +57 7

5. The proposed scheme o Accurate gradient selection predictor (AGSP) f g h e c b i d a x Causal template 8

5. The proposed scheme (cont. ) Example 1 : 45 40 55 50 Dh=10, Dv=30, D+=29, D-=35 50 65 54 Ch=105, Cv=65, C+=54, C-=50 102 105 100 =(10*54 + 29*105)/(10+29)=92 e = +8 Example 2 : 55 60 60 60 100 50 45 50 55 100 Dh=19, Dv=27, D+=21, D-=8 Ch=55, Cv=50, C+=45, C-=100 =(8*55 + 19*100)/(8+19)=87 e = +13 9

6. Experimental results o Test images: gray scale, 512 × 512 LOCO-I CALIC AGSP Amplitude images for prediction errors 10

6. Experimental results (cont. ) o Compression performance 11

7. Conclusions o o A new adaptive prediction algorithm based on accurate gradient estimation and selection All the possible contexts are considered in context modeling Handles complex structures more robustly Maintain the simplicity of implementation and computation 12