Hierarchical partitions for content image retrieval from large-scale database

Abstract

Increasing of multimedia applications in commerce, biometrics, science, entertainments etc. leads to a great need of processing of digital visual content stored in very large databases. Many systems combine visual features and metadata analysis to solve the semantic gap between low-level visual features and high-level human concept, i.e. there arises a great interest in contentbased image retrieval (CBIR) systems. As retrieval is computationally expensive, one of the most challenging moments in CBIR is minimizing of the retrieval process time. Widespread clustering techniques allow to group similar images in terms of their features proximity. The number of matches can be greatly reduced, but there is no guarantee that the global optimum solution is obtained. We propose a novel hierarchical clustering of image collections with objective function encompassing goals to number of matches at a search stage. Offered method enables construction of image retrieval systems with minimal query time. © Springer-Verlag Berlin Heidelberg 2005.