Image Retrieval Using Haar Discrete Wavelet Transform Advisors: Chi-Hong Lin, Yung-Kuan Chan Speaker: Chin-Hong Lin ( 林林林 ) Date: 2005-05-04
Jan 27, 2016
Image Retrieval Using Haar Discrete Wavelet Transform
Advisors: Chi-Hong Lin, Yung-Kuan Chan
Speaker: Chin-Hong Lin (林志鴻 )
Date: 2005-05-04
Outline
• Introduction• Feature of image• Image Retrieval• Experiments• Conclusion
Introduction
LL1 HL1
LH1 HH1
LL2 HL2
LH2 HH2
LL3 HL3
LH3 HH3
LEVEL1LEVEL2LEVEL3
LL:低頻
HL 、 LH:中頻
HH:高頻
Haar Discrete Wavelet Transform
Level 3
Feature of image
N
iixN
Mean1
1
Mean
Std)deviation( Standard
(Sk) Skewness
N
ii Meanx
NStd
1
2)(1
1
N
i
i
Std
Meanx
NSk
1
3)(1
Each Block for R 、 G 、 B Images
N : total pixels of each block
LEVEL 3
Feature of image
• Image features for R Image
RKRK
RK
RRRRRR SkStdMeanSkStdMeanSkStdMean ,,,...,,,,,, 222111
個90)skewness deviation, standard mean,(3)block(01)BG,R,(3
When LEVEL=3 (K 個 block)
• Image features
個30)skewness deviation, standard mean,(3)block(01
Image Retrieval System
• Distance between query and database images
K
i
rBi
Gi
RiDist
1
1
r
Ri
Ri
Ri
RiRSk
i
r
Ri
Ri
Ri
RiRStd
i
r
Ri
Ri
Ri
RiRMean
iRi
mSkMSk
SkSkW
mStdMStd
StdStdW
mMeanMMean
MeanMeanW
r
Gi
Gi
Gi
GiGSk
i
r
Gi
Gi
Gi
GiGStd
i
r
Gi
Gi
Gi
GiGMean
iGi
mSkMSk
SkSkW
mStdMStd
StdStdW
mMeanMMean
MeanMeanW
r
Bi
Bi
Bi
BiBSk
i
r
Bi
Bi
Bi
BiBStd
i
r
Bi
Bi
Bi
BiBMean
iBi
mSkMSk
SkSkW
mStdMStd
StdStdW
mMeanMMean
MeanMeanW
Image Retrieval System
• Weight (principal component analysis)
WWS
其中
kkkkkk
k
k
SSS
SSS
SSS
S
21
22221
11211
RMeank
RMean2
RMean1
W
W
W
W
n
lMean
RjlMean
Rilij ji
MeanMeann
S1
))((1
1
Experiments
• 系統環境
Database Images: 500張 Query Image: 500張
• Haar Discrete Wavelet Transform LEVEL=6
Experiments (Weight for R)
1 2
5
83 4
6 7
9 10
i
1 0.191913 0.245016 0.236156
2 0.091631 0.194320 0.060509
3 0.120890 0.000036 0.044046
4 0.026260 0.000372 0.064574
5 0.187342 0.100395 0.068993
6 0.015310 0.106159 0.168741
7 0.011028 0.085508 0.002203
8 0.272071 0.002706 0.243292
9 0.000147 0.166620 0.001367
10 0.083409 0.098867 0.110119
RMeaniW
RStdiW
RSkiW
R IMAGE
Experiments (Weight for G)
1 2
5
83 4
6 7
9 10
i
1 0.175647 0.321353 0.391986
2 0.014821 0.104429 0.000194
3 0.231864 0.004607 0.143521
4 0.012855 0.000525 0.002009
5 0.195031 0.150847 0.256818
6 0.005175 0.015622 0.004402
7 0.022285 0.145268 0.027417
8 0.251616 0.001902 0.100477
9 0.002430 0.156640 0.004652
10 0.088277 0.098806 0.068524
GMeaniW
GStdiW
GSkiW
G IMAGE
Experiments (Weight for B)
1 2
5
83 4
6 7
9 10
i
1 0.107564 0.091958 0.421789
2 0.124100 0.317156 0.011024
3 0.168880 0.000112 0.002285
4 0.022146 0.009671 0.132831
5 0.187198 0.190587 0.090834
6 0.007131 0.000868 0.029428
7 0.016765 0.109951 0.005432
8 0.244408 0.000215 0.237243
9 0.007249 0.181187 0.000182
10 0.114560 0.098296 0.068951
BMeaniW
BStdiW
BSkiW
B IMAGE
Experiments
• 查詢結果
ACC 方法 (%) L
HDWT CVAAO CCH FCH
1 94.2 91.8 80.8 83.4
2 96.6 94.8 87.6 87.4
3 97.6 96.8 90.8 89.4
4 97.8 97 92.2 90.2
5 98.2 97.4 93.6 92.2
Experiments
Rotation variant images Shift variant images
Experiments
Texture variant images
Noise variant images
Noise variant images(indecipherable )
Conclusion
• 本論文利用頻帶之間的權重突顯出影像特性達到加強查詢效果。
• 本論文實驗結果能提供不錯的查詢效果,且對影像大小變異、物件位移變異、旋轉變異以及紋理變異,均能有不錯的抵抗能力。
THE END THANK YOU