A New High-Dimensional Index Structure Using a Cell-Based Filtering Technique

Abstract

In general, multimedia database applications require to support similarity search for content-based retrieval on multimedia data, i.e., image, animation, video, and audio. Since the similarity of two multimedia objects is measured as the distance between their feature vectors, the similarity search corresponds to a search for the nearest neighbors in the feature vector space. In this paper, we propose a new high-dimensional indexing scheme using a cellbased filtering technique which supports the nearest neighbor search efficiently. Our Cell-Based Filtering (CBF) scheme divides a high-dimensional feature vector space into cells, like VA-file. However, in order to make a better effect on filtering, our CBF scheme performs additional filtering based on a distance between an object feature vector and the center of a cell including it, in addition to filtering based on cell signatures before accessing a data file. From our experiment using high-dimensional feature vectors, we show that our CBF scheme achieves better performance on the nearest neighbor search than its competitors, such as VA-File and X-tree.