Feature representations for image retrieval: beyond the color histogram

We study solutions to the problem of feature representation in the context of content-based image retrieval (CBIR). Retrieval is formulated as a classification problem, where the goal is to minimize probability of retrieval error. Under this formulation, retrieval performance is directly related to the quality of density estimation which is, in turn, determined by properties of the feature representation. We show that most representations of interest for the retrieval problem are particular cases of the mixture model, and present detailed arguments for why this is the most appropriate representation for retrieval