Wavelet-based color texture retrieval using the independent component color space

In this paper, we propose a wavelet based color texture retrieval method using the independent component color space. In color texture retrieval, the product of low dimensional marginal distributions of wavelet coefficients from different color layers are preferred to substitute or approximate their high dimensional joint distributions in order to avoid the curse of dimensionality. However, the RGB color spaces is a highly correlated color space and the extractedwavelet coefficients from different layers are also correlated, which means such a substitution or approximation will not be adequate. To solve the problem, we use independent component analysis to decorrelate the R, G and B layers into three new independent layers before applying wavelet decomposition on the color texture images. In the feature extraction(FE) step of the proposed method, generalized Gaussian density (GGD) are used to model the marginal distribution of wavelet coefficients, and the extracted model parameters are used as features. In the similarity measurement (SM) step of the proposed method, the Kullback-Leibler distance(KLD) is calculated as feature distance, using the extracted model parameters of the query texture images and those of the images in the database. Experimental results on a database of 1120 color texture images indicate that the proposed method greatly overperforms its RGB based counterpart that ignores the inter-layer correlation, and its counterpart which uses the I1I2I3 colorspace.