Research On Optimized Residual Vector Quantization For Nearest Neighbor Search In Image Retrieval

With the rapid development of digital image technology,image-based applications have become more widespread,therefore image retrieval has become an important research topic in the field of computer vision.In practical applications,the approximate nearest neighbor search is proposed in order to reduce the search consumption.Approximate nearest neighbor search does not require the most accurate results,but finds results with the largest probability.Vector quantization is an efficient compact coding technique that is commonly used in approximate nearest neighbor search.By quantizing and reducing the dimension of the high-dimensional feature vectors,the computational cost of the search process can be effectively reduced.However,the current vector quantization method has the problem of excessive information loss,and the final approximate nearest neighbor search result is not satisfactory.This thesis improves the existing vector quantization methods and proposes the topic of parametric and nonparametric residual vector quantization.Based on residual vector quantization which is a classical method of vector quantization,this thesis uses a multi-stage quantization strategy,which takes residual vector of last stage as the input of current quantizer,so that the quantization distortion could be reduced.According to statistical learning,nonparametric and parametric optimization methods are proposed to improve the data space distribution and codebook of each stage of quantization,which further reduces the vector quantization error and is helpful to obtain better quantization results.To validate the effectiveness of the proposed method,experiments and analyses are conducted on two benchmark dataset,and it is proved that the proposed method can achieve satisfactory results with good stability and robustness.