Codewords-Expanded Enhanced Residual Vector Quantization for Approximate Nearest Neighbor Search

Abstract

A codewords-expanded enhanced residual vector quantization (CERVQ) is proposed to improve the accuracy of approximate nearest neighbor (ANN) search for feature vectors. It combines enhanced residual vector quantization (ERVQ) with the method of calculating mean-equisection vector to reduce training error and improves the quantization accuracy. Firstly, the mean-equisection vector is used to compute residual vector as the input to next layer in the training stage, except for the first layer. Based on this, an iterative method for optimizing codebooks is designed to reduce overall quantization error. Secondly, the codebook of each layer is expanded with the mean-equisection vectors to generate new codewords in quantization stage, which are used to quantize input feature vectors to improve quantization accuracy. Finally, a method of calculating asymmetric Euclidean distance is implemented for ANN search. The CERVQ is compared with 5 typical methods on two public SIFT and GIST datasets. Experiment results show that the training error reduces by 10%~24% and the recall rate of ANN search is increased by 1%~44%. In addition, the scale of codebook in the CERVQ can reduce by 50% under the condition of reaching the same recall rate.