Research On Key Technology Of Quantization Based Image Feature Retrieval

The application of large-scale image retrieval has became more and more popular.Due to the advantages of low precision loss and high computational efficiency,the approximate nearest neighborhood(ANN)retrieval algorithm has received more and more attention in the field of large-scale image retrieval.In many ANN algorithms,quantization retrieval algorithm with high precision has became popular in the field of academic research.This paper mainly focuses on the application of reinforcement learning algorithm based image quantization retrieval.The gradient strategy in reinforcement learning is used to solve the codebook optimization problem in quantization retrieval and the traditional retrieval methods.They are integrated to improve the effectiveness of the quantization retrieval process in this paper.The main contributions of this paper are as followes:(1)A new iterative quantization based shared codebook retrieval algorithm is proposed in this paper.We propose a iterative quantization algorithm based on shared codebook,which not only improves the quantization coding speed and optimizes the complexity of the algorithm,but also reduces the searching process of the code of the storage space greatly.(2)An optimization method based on reinforcement learning is proposed to solve codebook update problem.This algorithm transforms coding process into a codewords sequence selection problem,so that the coding optimization process can be integrated into the reinforcement learning algorithm.(3)An end-to-end supervised image retrieval system which integrates image feature extraction and quantization module is designed.Compared with feature-based Euclidean distance direct retrieval,this system has the advantage of the higher computational efficiency.Compared with hash retrieval methods,this system makes full use of the advantage of precision in quantization algorithm.