Information theoretic similarity measures for content based image retrieval

Content-based image retrieval is based on the idea of extracting visual features from image and using them to index images in a database. The comparisons that determine similarity between images depend on the representations of the features and the definition of appropriate distance function. Most of the research literature uses vectors as the predominate representation given the rich theory of vector spaces. While vectors are an extremely useful representation, their use in large databases may be prohibitive given their usually large dimensions and similarity functions. In this paper, we propose similarity measures and an indexing algorithm based on information theory that permits an image to be represented as a single number. When use in conjunction with vectors, our method displays improved efficiency when querying large databases.