Database retrieval: the use of combined dissimilarities

In image retrieval systems, the key point is the description of the set of images. In this paper we show that a representation using a cloud of points offers a flexible description but suffers from class overlap. We propose a novel approach for describing clouds of points based on the support vector data description (SVDD). We show that combining image descriptions using dissimilarities improves the retrieval precision. Further we propose a method to select an efficient and robust subset of classifiers. We investigate the performance of the proposed retrieval technique on a database of 368 images.