Hierarchical Indexing Scheme for Fast Search in Large-Scale Image Database

Hangjun J. Ye,Guangyou Xu

DOI: https://doi.org/10.1117/12.539947

2003-01-01

Abstract:Practical content-based image retrieval systems require efficient indexing schemes for fast k-nearest neighbor (k-NN) searches. Researchers have proposed many tree-based methods using space and data partitioning for similarity searches. However, traditional indexing methods perform poorly and will degrade to simple sequential scans at high dimensionality - that is so-called "curse of dimensionality". Recently, several filtering approaches based on vector approximation (VA) were proposed and showed promising performance. However, VA-based approaches need compute the bound of the distance between each feature vector and the query. It will consume the same computational overhead as the brute-force sequential scan. In this paper, a novel hierarchical indexing scheme is proposed. This approach integrates VA-based index structure with approximate NN (ANN) searches and performs probabilistic ANN searches on approximate vectors. Experiments show the proposed approach achieves a remarkable reduction of computational overhead and disk accesses for k-NN searches. This presented approach supports quadratic-form distance metric and can integrate with relevance feedback techniques for practical large-scale image retrieval systems.