Fast vector quantization image coding by mean value predictive algorithm

Fast vector quantization

image coding by mean value

predictive algorithm

Authors: Yung-Gi Wu, Kuo-Lun FanSource: Journal of Electronic Imaging 13(2), 324–333 (April2004).

Speaker: Meng-Jing TsaiDate: 2012.05.08

2

Outline• Introductions– Vector Quantization (VQ)– Accelerate Coding Techniques• Partial Distance Search (PDS) Algorithm• Fast Euclidean Computation Technique

• Proposed Coding Scheme– Preprocessing process– Practical Encoding Process

• Simulation Results• Conclusions

3

Training set

.

.

.

01...

N-2N-1

Introduction - VQ

• LBG Algorithm

Initial codebook

01...

254255

Training Images

4

Introduction - VQ

• Image encoding procedure

ImageIndex table

wh

5

Introduction - VQ

• Image decoding procedure

ImageIndex table

wh

6

Introduction – VQ & PDS

• The traditional VQ compression system must search from the first codeword to the last one.

• It costs a lot of time.• In order to accelerate the computation, the

PDS algorithm provides an effective coding technique.

7

c10 c11 c12 c13 c115

c00 c01 c02 c03 c015

Introduction - PDS

.

.

.

012

.

.

.

253254255

c20 c21 c22 c23 c215

v0 v1 v2 v3

v4 v5 v6 v7

v8 v9 v10 v11

v12 v13 v14 v15

Codebook

Vector

＝ dmind > dmin

＝ dmin

d < dmin

8

Introduction

• Fast Euclidean Computation Technique Look-up Table (LUT)

9

Proposed Coding Scheme

• Two parts Preprocessing process

1. A codebook that is sorted by the mean value of each codeword within the codebook.

2. The sorted mean value table of each codeword.

3. The 2D fast Euclidean table. Practical encoding process

Proposed Coding Scheme• Pratical Encoding Process

50 80 155 220

100 70 80 210

85 60 85 190

105 90 85 75

mean = 108.75

06...

98108112

.

.

.219233

1st codeword

2nd codeword

.

.

.(FSP-1)th codeword

First Search Point (FSP) codeword

(FSP+1)th codeword

.

.

.254th codeword

255th codewordmean value table codebook 10

11

Practical Encoding Process

PSNR = 20.499 dB

12

Pratical Encoding Process

The searching direction.

13

Simulation Results

• First situation– codebook1 & Lenna

• Second situation– codebook2 & Jet, Pepper, Lenna, and Scene

14

Simulation Results

• First situation

PSNR and window size (codebook1 and 2D fast Euclidean technique)

PSNR and window size (codebook1 and 1D fast Euclidean technique)

Simulation Results

• First situation

15

16

Simulation Results

• First situation

Simulation Results

• First situationThe average arithmetic operation for each pixel needed with the codebook1.

17

18

Simulation Results

• Second situation– codebook2 and 2D fast Euclidean technique

Window size and CPU time Window size and PSNR

19

Simulation Results

• Second situationThe average arithmetic operation for each pixel needed with the codebook2.

20

Conclusions

• The PDS algorithm and the (2D) fast Euclidean technique are used to accelerate the calculation.

• It needs more memory for extra tables.• Compared to a full search algorithm, it saves

more than 95% searching time.

Thank you for your listening.