An efficient nearest neighbor searching algorithm with application to LBG codebook gene ration

Abstract

The nearest neighbor (NN) searching problem has wide applications. In vector quantization (VQ), both the codebook generation phase and encoding phase (using the codebook just generated) often need to use the NN search. Improper design of the searching algorithm will make the complexity quite big as vector dimensionality k or codebook size N increases. In this paper, a fast NN searching method is proposed, which can then accelerate the LEG codebook generation process for VQ design. The method successfully modifies and improves the LAESA method. Unlike LAESA, the proposed k/2 ''fixed'' points (allocated far from the data) and the origin are used as the k/2+1 reference points to reduce the searching area. The overhead in memory is only linearly proportional to N and k. The time complexity, including the overhead, is of order O(kN). According to our experiments, the proposed algorithm can reduce the time burden while the distortion remains identical to that of the full search.