Indexing Images in High-Dimensional and Dynamic-Weighted Feature Spaces

As information retrieval systems evolve to deal with multimedia content, we see the dimensions of content feature space increasing, and relevance feedback being employed to provide more accurate query results. In this paper, we propose using Tree-structured Vector Quantization (TSVQ) to index high-dimensional data for supporting efficient similarity searches and effective relevance feedback. To support efficient similarity searches, we first use TSVQ to cluster data and store each cluster in a sequential file. We then model a similarity search as a classification problem—similar objects are much more likely to be found in the clusters into which the query object is classified. When relevance feedback is considered, and thereby features are weighted differently, we show that our approach remains very effective. Our empirical study on both a 51K and a one-million-image dataset shows that tackling indexing as a classification problem and solving the problem with TSVQ is efficient, effective, and scalable with respect to both data dimensions and dataset size.