Minimum Probability of Error Image Retrieval: From Visual Features to Image Semantics

The recent availability of massive amounts of imagery, both at home and on the Internet, has generated substantial interest in systems for automated image search and retrieval. In this work, we review a principle for the design of such systems, which formulates the retrieval problem as one of decision-theory. Under this principle, a retrieval system searches the images that are likely to satisfy the query with minimum probability of error (MPE). It is shown how the MPE principle can be used to design optimal solutions for practical retrieval problems. This involves a characterization of the fundamental performance bounds of the MPE retrieval architecture, and the use of these bounds to derive optimal components for retrieval systems. These components include a feature space where images are represented, density estimation methods to produce this representation, and the similarity function to be used for image matching. It is also Full text available at: http://dx.doi.org/10.1561/2000000015