Integrating Color And Spatial Information for CBIR NTUT CSIE D.W. Lin
References L. Cinque, G. Ciocca, S. Levialdi, A. Pellicano, and R. Schettini, “Color-based image retrieval using spatial- chromatic histograms,” Image and Vision Computing, 19 (2001) M.S. Kankanhalli, B.M. Mehtre, and H.Y. Huang, “Color and spatial feature for content-based image retrieval,” Pattern Recognition Letters, 20 (1999) S. Berretti, A.D. Bimbo, and E. Vicario, “Spatial arrangement of color in retrieval by visual similarity,” Pattern Recognition 35(2002)
The representation of color Color histogram – Global color histogram – Global color histogram + spatial info. – (fixed) Partition + local color histogram Dominant color – Extracting the representative colors of image via VQ or clustering (e.g. k-means algorithm) – Spatial info. can be attained Histogram refinement Specific-color pixel distribution (single, pair, triple …) Edge histogram … Non-adaptive
Spatial-chromatic histograms [1] SCH – global color histogram with info. about (single) pixel distribution SCH attempts to answer: – How many meaningful colors? color space quantization – Where the pixels having the same color? location of region(with same color) – How are these pixels spatially arranged? distribution of region
SCH – Color representation Color representation – CIELAB Munsell ISCC-NBS CIELAB – CIEXYZ CIELAB CIELUV – a, b: opponent color ( green red, blue yellow )
SCH – Color representation Munsell color system – Hue value chroma
SCH – Color representation ISCC-NBS Centroid Color System – Partitioning the Munsell color system into 267 blocks, each blocks represented by an unique linguistic tag and the block centroid (Munsell coordinates) Using back-propagation NN to transform – CIELAB Munsell ISCC-NBS
SCH – Feature vector The definition of SCH for image I S I (k) = (h I (k), b I (k), σ I (k)) – k: k th quantized color (1~c) – h I (k): pixel amount(ratio) – b I (k): baricenter (normalized mean coordinates) – σ I (k): standard deviation of (spread) Properties: – Insensitive to scale changes(via normalization) – Compact representation and rapidly computing
SCH – Similarity measure Similarity function – c: number of quantized color – d(·): Euclidean distance – : max. distance
SCH – Effectiveness measure – S: relevant items in DB – : retrieved set (short list) for a query – : relevant items in retrieved set
Color and spatial clustering [2] k-means algorithm – Iteration version – Two-pass version – VQ (LBG algorithm) Proposed color clustering (two-pass) – Generating a new cluster while d(p, C i ) > T – Merging those clusters with small population to the nearest cluster
Color and spatial clustering Spatial clustering – Based on the clustered color layer – Using connected components labeling to separate the spatial clusters – Discarding those clusters with small population or lower density(embedded rectangular)
Feature vectors For image I, color clusters can be given C ci = {R i, G i, B i, λ ci, x ci, y ci } i: 1..m (number of color cluster) R i, G i, B i : representative color of cluster λ ci : pixel ratio of cluster to total x ci, y ci : centroid of cluster f c ={C ci |i=1, 2, …, m} Do the same to color-spatial clusters
Similarity measure – 1 : color distance between color cluster(RGB) – 2 : relative frequency of pixels of color cluster ( ) – 3 : spatial distance between color cluster(x, y) – 4 : relative frequency of pixels of color-spatial cluster ( ) – 5 : spatial distance between color-spatial cluster(x, y)
Spatial arrangement of color[3] The back-projection from dominant colors to the image results in an exceedingly complex model(e.g. [2]) Authors proposed a descriptor, called weighted walkthroughs, to capture the binary directional relationship of two complex sets of pixels
Weighted walkthroughs – The model can be extended to represent the relationship of two sets A, B – w 11 evaluates the number of pixel pairs a A and b B such that b is upper right from a +,+ w +1+1 a B +, - -, + -, - C i : characteristic function of negative/positive real number |B| : area of Bi,j : ±1
Compositional computation Reducing the region to a set of rectangular The weight between A and B 1 B 2 can be derived by linear combination of A/B 1 and A/B 2
Distance of WW 3 directional indexes: – w H (A, B) = w 1, 1 (A, B) + w 1, -1 (A, B) – w V (A, B) = w -1, 1 (A, B) + w 1, -1 (A, B) – w D (A, B) = w -1, -1 (A, B) + w 1, 1 (A, B) Spatial distance A B A B +,+ w +1+1 a B +, - -, + -, - A B
Arrangements comparing Image model: E: set of spatial entities (color-clustered region) a: E A { any a } (chromatic label) w: E E W { any s } (spatial description)
Arrangements comparing Distance between image model Q and D – : injective function(interpretation, association between query and model image) – D A : chromatic distance (L*u*v*) – D S : spatial distance – N q : number of entities in query
Future works Finish the color-spatial study(geometric- enhanced histogram) Study wavelet and JPEG 2000