Content Based Image Retrieval by combining color, texture and CENTRIST

This paper presents a novel framework for Content Based Image Retrieval(CBIR), which combines color, texture and spatial structure of image. The proposed method uses color, texture and spatial structure descriptors to form a feature vector. Images are segmented into regions to extract local color, texture and CENTRIST(CENsus Transform hISTogram) features respectively. Multiple-instance learning (MIL) and Diverse Density(DD) are incorporated with regions as instances to find the objective instance. In addition, to denote the whole structure of image better, we perform PCA to CENTRIST features of all images, i.e. spatial Principal component Analysis of Census Transform(spatial PACT). This framework integrates three features to enhance the retrieval performance. Experiments on COREL standard database invalidate the proposed method by comparing with some state-of-the-art methods. (4 pages)