Abstract: Many traditional indexing methods perform poorly in the high-dimensional vector space. The Vector Approximation File approach overcomes some of the difficulties of curse of dimensionality. A new k-nearest neighbor search algorithm based on VA-File for relevance feedback image retrieval is introduced in this paper. Based on the feedback, the correlations of the underlying similarity metric between two consecutive searches is exploited, and then the search result and feedbacks and used to filter the approximate vectors in the next search round. Experiments on the large real-world dataset show a remarkable reduction of vectors accessed and an improvement on the indexing performance compared with the existing search algorithm.

Key words: CBIR(content-based image retrieval), high-dimensional indexing, relevance feedback, vector approximation, k-nearest neighbor search