Post on 05-Feb-2016
description
Image indexing and Retrieval Using Histogram Based Methods,
03/6/503/6/5
資工研一資工研一陳慶鋒陳慶鋒
Outline
Histogram based methodsHistogram based methods ImplementationImplementation Experiment resultExperiment result Future workFuture work ReferencesReferences
General formula in successful IR
A feature vector A feature vector f(I)f(I) for image for image II II and and I’I’ are not “similar” are not “similar”
if and only if if and only if |f(I)-f(I’)||f(I)-f(I’)| is large is large f(.) should be fast to computef(.) should be fast to compute f(I) should be small in sizef(I) should be small in size
Color histogram
For a For a nnnn with with mm colors image colors image II,,
the color histogram is the color histogram is
wherewhere
pp 為屬於為屬於 II 的的 pixel, pixel, I(p)I(p) 為其顏色為其顏色 , ,, ,forfor
Color histogram (cont.)
Distance measure:Distance measure:
令原圖為令原圖為 II ,,欲比對的圖為欲比對的圖為 I’I’
在比對上使用在比對上使用 LL11-distance -distance ::
比對方式:比對方式: thresholdingthresholding
Color histogram (cont.2)
AdvantagesAdvantages
-trivial to compute-trivial to compute
-robust against small changes in camera -robust against small changes in camera
viewpointviewpoint DisadvantagesDisadvantages
-without any spatial information-without any spatial information
Histogram refinement
The pixels of a given bucket are subdivided The pixels of a given bucket are subdivided into classes based on local feature. Within a into classes based on local feature. Within a given bucket , only pixels in the same class given bucket , only pixels in the same class are compared.are compared.
The local feature which this paper used:The local feature which this paper used:
Color Coherence Vectors(CCVs)Color Coherence Vectors(CCVs)
Histogram refinement (cont.)
CCVsCCVs
For the discretized color For the discretized color jj, the pixels with color , the pixels with color jj are coherence if they are adjacent(using eight-are coherence if they are adjacent(using eight-neighbor), indicated as neighbor), indicated as jj, otherwise are , otherwise are
incoherence, indicated as incoherence, indicated as jj, and total pixel with , and total pixel with
color color jj= = jj+ + jj, , a threshold a threshold is defined as the is defined as the
condition of coherence or notcondition of coherence or not
for color for color jj, the coherence pair is (, the coherence pair is (jj, , jj) )
Histogram refinement (cont.2)
CCVs (cont.)CCVs (cont.) Comparing CCV with L1 distance:Comparing CCV with L1 distance:
Distance measure:Distance measure:
比對方式: 比對方式: thresholdingthresholding
Histogram refinement (cont.3)
ExtensionExtension
Centering refinementCentering refinement
Successive refinementSuccessive refinement
Color correlograms
A new image featureA new image feature Robust against large changes in camera Robust against large changes in camera
viewpointviewpoint
Color correlograms (cont.)
A table indexed by color pairs, where the A table indexed by color pairs, where the kk-th entry for -th entry for color pair color pair <i, j><i, j> specifies the probability of finding a pixel specifies the probability of finding a pixel of color of color jj at a distance at a distance kk from a pixel of color from a pixel of color ii in the image. in the image.
The correlogram isThe correlogram is
The autocorrelogram is The autocorrelogram is
Color correlograms (cont.2)
Properties:Properties:
-Contains spatial correlation of colors-Contains spatial correlation of colors
-Easy to compute-Easy to compute
-The size of feature is fairly small (-The size of feature is fairly small (O(md)O(md)))
Implementation
PreprocessPreprocess Sizes of all images are normalized to 192*128Sizes of all images are normalized to 192*128
Colors of all images are quantized to 16Colors of all images are quantized to 16
Set Set of CCV as 2500 of CCV as 2500
Set Set d d of autocorrelogram as 30of autocorrelogram as 30
Implementation(cont.)
IndexingIndexing
color histogramcolor histogram
CCVCCV
Implementation(cont.)
Indexing(cont.)Indexing(cont.) color autocorrelogramcolor autocorrelogram
Implementation(cont.)
Similarity measureSimilarity measure
Experiment result
Sample queries and answers with ranks for Sample queries and answers with ranks for various methodsvarious methods
hist:2 ccv:1 auto:3hist:2 ccv:1 auto:3
Experiment result(cont.)
hist:12 ccv:11 auto:4hist:12 ccv:11 auto:4
hist:29 ccv:24 auto:15hist:29 ccv:24 auto:15
Experiment result(cont.)
hist:8 ccv:9 auto:18hist:8 ccv:9 auto:18
hist:7 ccv:23 auto:15hist:7 ccv:23 auto:15
Future work
Use color imagesUse color images Study more about tech of CBIRStudy more about tech of CBIR
References
[1][1] G. Pass and R.Zabih, “histogram refinement for content G. Pass and R.Zabih, “histogram refinement for content based image retrieval,” IEEE Workshop on Applications based image retrieval,” IEEE Workshop on Applications of Computer Vision, pp.96-102, 1996of Computer Vision, pp.96-102, 1996
[2] J. Huang, S. R. Kumar, M. Mitra, W. J. Zhu, and R.Zabih, [2] J. Huang, S. R. Kumar, M. Mitra, W. J. Zhu, and R.Zabih, “Image indexing using color correlograms,” Conf. “Image indexing using color correlograms,” Conf. Computer Vision and Pattern Recognit., pp.762-768,1997Computer Vision and Pattern Recognit., pp.762-768,1997
[3]G. Pass, R. Zabih, and J. Miller, “Comparing images using [3]G. Pass, R. Zabih, and J. Miller, “Comparing images using color coherence vectors”, color coherence vectors”, Proc. of ACM MultimediaProc. of ACM Multimedia 96, 96, pp. 65-73, Boston MA USA, 1996pp. 65-73, Boston MA USA, 1996