A fast algorithm of k-nearest neighbor search in the similarity retrieval of image databases

In this paper, we propose a fast algorithm for k-nearest neighbor search in high dimensions. The complexity of most existing algorithms including k-d tree and R-tree grows exponentially with dimension. In content-based image retrieval, the number of dimensions of feature vectors is more than 102. Static space partitioning methods are not suitable to relevance feedback for realizing human in the loop. Moreover, the performance of the projection algorithm proposed by Friedman et. al which is well known as the basic method of dynamic space partitioning methods degrades with increasing dimentionarity. In this paper, we propose an approximation of the projection algorithm to improve the performance in high dimensionality. The proposed algorithm is verified by simulation in which the correctness and computation time of both algorithms are compared. In simulation, both synthesized and real image data set are used. Our algorithm obtains at least 90% correctness in 50% search time relative to that by the exhaustive search in 200 dimensional space.