adaptive edge-based side-match finite-state classified vector quantization with quadtree map
DESCRIPTION
Adaptive Edge-Based Side-Match Finite-State Classified Vector Quantization with Quadtree Map. IEEE transactions on image processing. VOL. 5, NO. 2, FREBRARY 1996 Authors Ruey-Feng Chang( 張瑞峰 ), CS, CCU Wei-Ming Chen( 陳偉銘 ), CS, CCU. Outline. Introduction of Vector Quantization (VQ) - PowerPoint PPT PresentationTRANSCRIPT
Adaptive Edge-Based Side-Match Finite-State Classified Vector Quantization with Quadtree Map
IEEE transactions on image processing.VOL. 5, NO. 2, FREBRARY 1996
AuthorsRuey-Feng Chang(張瑞峰 ), CS, CCUWei-Ming Chen(陳偉銘 ), CS, CCU
Outline
Introduction of Vector Quantization (VQ) Basic VQ techniques Adaptive edge-based side-match Simulation results Conclusion
Introduction of Vector Quantization (1/2) Efficient scheme for image compression Component
Codebooks Generated by using the iterative clustering algorithm
Encoder Image is first partitioned into non-overlapping rectangular
blocks (vectors) Each vector is quantized (indexed) to the closest codeword in
the codebook Decoder
Select the corresponding codeword in the codebook via indexes
Encoder side
Decoder side
Clustering
algorithmCodebook
Yi, i = 1, …, Nc
Training image set
Target image
Partition image to NxN blocks (vectors)
Find the closet codeword and index for each vector
Re-constructing image
Find the corresponding codeword via indexes
Introduction of Vector Quantization (2/2) What is closest codeword
Small Euclidean distance
How to generate codebooks Cluster algorithm
K-means Linde-Buzo-Gray (LBG) …
Basic VQ techniquesClassified Vector Quantization (CVQ)
Features Multiple codebooks for specified features of
blocks Advantage
Reduce search time Disadvantage
Extra bits needed
Basic VQ techniquesFinite-State Vector Quantization (FSVQ)
Features Similar to CVQ, but the used codebook is decided
by current codebook and current codeword Advantage
Reduce search space Extra bits aren’t needed
Disadvantage Derailment
Basic VQ techniquesSide-Match Vector Quantization (SMVQ) (1/2)
Features A class of FSVQ, but use the side of upper and
left neighboring blocks to generate the state codebook
Advantage Reduce search space Smoother
Disadvantage Derailment
Basic VQ techniquesSide-Match Vector Quantization (SMVQ) (2/2)
n
jmjj uyyhd
1
21 )()(
m
iini lyyvd
1
21 )()(
)()()( yvdyhdysmd
Adaptive edge-based side-match
Edge Detection Sobel Filter
Classification Non Edge Block (SMVQ) Edge Block (CVQ,SMVQ)
Adaptive edge-based side-matchSobel Filter (1/2)
Sobel Filter can increase the high frequency part of image.
Formula
Gradient :
Θ : threshold for checking if the edge occur
|||| yx GGf
||||,0
||||,1
GyGx
GyGxEage
Adaptive edge-based side-matchSobel Filter (2/2)
Sobel operator Gy = (z3 + 2z6 + z9) – (z1 + 2z4 + z7)
Gx = (z7 + 2z8 + z9) – (z1 + 2z2 + z3)
101
202
101
121
000
121
987
654
321
zzz
zzz
zzz
Image region Mask used to compute Gx Mask used to compute Gy
Adaptive edge-based side-matchQuadtree Map
QTC=1-0011-0001-0011
Problem Each vector need one more bit to determine its class
For example, in a 512 X 512 image with 4 X 4 block size, 16386 bits must be transmitted to the decoder
Solution Quadtree map
Adaptive edge-based side-matchEdge Block(CVQ)
Nonedge blocks encode first. Original SMVQ
The edge blocks are classified into 16 subclasses, according to the neighboring blocks which are edge or nonedge.
Adaptive edge-based side-matchEncode
Adaptive edge-based side-matchDecode
Simulation results (1/5)
Contribution Higher quality with the same bit rate Codebooks size are variable
Test arguments 256 gray level image Image size : 512 x 512 Vector size : 4 x 4
dBMSE
PSNR2
10
255log10
Test criterion
m
i
m
jijij xx
mMSE
1 1
22 )ˆ()1(
Simulation results (2/5)
Simulation results (3/5)
Simulation results (4/5)
Simulation results (5/5)
Conclusion
The classified FSVQ combine the advantages of CVQ and SMVQ
The system complexity is higher