Incremental feature weight learning and its application to a shape-based query system

Abstract

Similarity between shapes is often measured by computing the distance between two feature vectors. Unfortunately, the feature space cannot always capture the notion of similarity in human perception. So, most current image retrieval systems use weights measuring the importance of each feature. However, the similarity does not vary with equal strength or in the same proportion in all directions in the feature space. In this paper, we present feature weights based on both clustered objects in the database and on relevance feedback. We show that using variance information from shape clusters to guide cluster information for an initial database search gives better results than using the standard Euclidean distance. To automatically incorporate a user's need, the proposed shape-based query system uses an incremental feature weight learning method that refines prototypes. In contrast to existing image database systems, the system can learn from user feedback. Indexing and retrieval results are presented that demonstrate the efficacy of our technique using the well-known Columbia database.