Image indexing and similarity retrieval based on key objects

Recently, many applications increasingly demand content-based image retrieval. To speed up the search process, a signature file containing the signatures associated with database images is frequently used as a filter to prune off non-promising images at the early stage of query processing. In this paper, we generate the signature for each picture with respect to "key" objects and propose a novel structure for organizing signatures. By using this new indexing structure, the number of signatures to be examined per query is reduced significantly. In particular, the number of bits to be examined per query is fewer as compared to the signature tree method proposed by Chen [1].