Content-based color image retrieval using multi-variate feature vectors

Abstract

This paper introduces a content-based image retrieval system which utilizes color, spatial frequency, and structural features of an object in the image. The major key color areas are extracted by image segmentation applying Bayesian classifier with k-means starter in CIELAB space and the color similarity is measured by summing up the inter-cluster color distances. The mutual correlations in DCT components are used to discriminate the spatial frequency features that carry textural details. In addition, a structural feature of image is simply characterized by down sampled color mosaic pattern. A target color image is retrieved by searching the image with the maximum similarity by cross correlations in the multi-dimensional feature vector space. Color similarity was indispensable to narrow the reliable candidate for almost all the tested images. Although an appropriate combination of spectral or structural features worked effective to retrieve the image with textural similarity, but we have not any definitive solution to select the frequency region yet. Finally, we report a method for evaluating the performance of our system based on psychophysical experiments using z-score.