A signature representation and indexing scheme of color-spatial information for similar image retrieval

Recently, many Web applications increasingly demand content based image retrieval. The authors present a two-axial signature based method of retrieving similar images based on color-spatial information for the design of an image search engine on the Internet. Two kinds of signatures, a dominant color composition (DCC) signature and a set of dominant color distribution (DCD) signatures are extracted from an image to represent the color-spatial information. In particular, DCD signatures are generated by projecting positions of each subregion with its own dominant color on the horizontal and vertical axis after partitioning an image area into n/spl times/n subregions. The DCC signature is used to prune away irrelevant images to a given image query, before evaluating the more expensive similarity measure on prospective candidate images. The spatial similarity of dominant colors is evaluated using the DCD signature. Experiments are performed on an image database of over 6000 images with our prototype system which adopts DCC and DCD signatures as index key for color-spatial image retrieval. Our results show the retrieval robustness for images that are partially rotated, transformed, and sub-sampled.