Research On Image Feature Retrieval Based On Multi-codebook Quantization

With the rapid development of multimedia and Internet technology,the number and growth rate of image data have reached a high level,and the speed and accuracy of image retrieval technology are facing huge challenges.In the existing image retrieval algorithms,hashing and quantization are widely used by the industry due to the low data storage space required.Among them,the accuracy of the quantization-based image retrieval algorithm is much higher than that of the hashing,which has huge research value and development prospects.This thesis focuses on the unsupervised image retrieval algorithm based on non-orthogonal multi-codebook quantization.First,the research process at home and abroad and the existing algorithms are sorted out and summarized.Then,using the AQ(Additive Quantization)model as an example,various problems encountered in non-orthogonal multi-codebook quantization and the corresponding solution algorithms are introduced in detail.Next,based on the state-of-the-art method LSQ(Local Search Quantization)in the AQ model,this paper first uses the C++ language and related optimization techniques to implement the algorithm,which improves the performance from the engineering aspect.Then this thesis analyzes the deficiencies of encoding process in LSQ from the algorithm level,proposes optimization modules including the vectorized calculation mode of tables,the parallelized calculation framework of encoding process,the enhancement of randomness and the new acceptance criteria.A large number of experiments on three datasets show the efficiency and effectiveness of these modules.The optimized algorithm is 2x faster than the original,and has higher retrieval accuracy and generalization.In addition,this thesis also discusses and optimizes the retrieval algorithm based on the AQ model,and gradually proposes three optimization algorithms: retrieval algorithm based on fast table calculation,non-exhaustive retrieval algorithm based on space partition,and retrieval optimization based on graph.Experimental results show that these algorithms can better weigh time,space and retrieval accuracy.Under certain circumstances,only a small amount of runtime memory is consumed,which can increase the retrieval speed by dozens of times while maintaining a high retrieval accuracy.