Research On Image Product Quantization Retrieval Algorithm For High Dimensional Features

With the rapid development of computer and Internet technology,both the scale and the dimension of multimedia data increases rapidly.Therefore,how to quickly and accurately retrieve target data from large-scale high-dimensional databases is a hot issue to be solved in the field of data mining.The Product Quantization(PQ)algorithm in Approximate Nearest Neighbor(ANN)search is widely applied due to its features such as small storage space,high search accuracy and easy implementation,but some problems exist such as the search speed of PQ needs to be further improved.Based on this background,a fast Product Quantization algorithm based on the Adaptive Subspace Dimension Product Quantization(ASDPQ)is proposed in this paper to solve the shortcomings of conventional PQ-based algorithms.The proposed ASDPQ is applied to Content-Based Image Retrieval(CBIR)system,and verified by this system.The research content of this paper is as follows:(1)A fast product quantization algorithm named Adaptive Subspace Dimension Product Quantization(ASDPQ).All of the approximate nearest neighbor search algorithms using the product quantization framework use a fixed number of subspaces.Since the amount of computational cost in calculating the total approximate distance is approximately proportional to the number of subspaces,optimizing the number of subspaces and guarantee the quantization accuracy can effectively reduce the amount of computation to retrieve more efficiently.Based on this,this paper proposes a fast PQ-based algorithm named Adaptive Subspace Dimension Product Quantization(ASDPQ)to accelerate the search,by reducing the average number of subspaces using adaptive selection of subspace dimensions.The experimental results conducted on SIFT and GIST datasets indicate that the ASDPQ algorithm proposed in this article can achieve a 33%acceleration effect as expected by the algorithm,while maintaining the same search accuracy.(2)Development of CBIR image retrieval System based on ASDPQ.In order to verify the effectiveness of the proposed algorithm,this paper develops an image retrieval system based on this algorithm,realizes the function of image retrieval by using ASDPQ,and displays the corresponding principle visually.In the process of system development,based on the commonly used image datasets,three kinds of feature datasets: SIFT,GIST and HSV are build,and aneasy-to-use and convenient user interface is designed.The experimental results show that the system can use multiple features for accurate and fast image retrieval in the datasets while provides a positive user experience.The effectiveness and practical application value of the proposed ASDPQ algorithm can be verified.