Research On Technology Of Content-Based Large-Scale Image Retrieval

With the popularization of the internet and digital imaging equipment, the online communities accumulated a variety of electronic images have been developing rapidly, the database containing a large number of images also continue to occur. How to retrieve quickly from the database to meet the needs of users has become an important theme. In order to achieve accurate and fast image retrieval, we carry out some detailed study from several aspects including image descriptor, large-scale approximate nearest neighbor search and indexing high-demensional vector.In view of the contradiction between Distinguishable image descriptor and storage of VLAD, an Enhenaced VLAD (EVLAD) is proposed. In formulation of EVLAD, image descriptor is generated by using of two level of visual codebook. The generation of residual vector of image local visual feature is based on the second level visual codebook, while the the accumulation of residual is based on the first level visual codebook. In addition, in view of uneven distribution of local visual feature, a measure is proposed to optimize the second level visual codebook. The experimental result shows that the stratege of two level codebooks and optimazation can improve image descriptor precision in the case of not increasing the size of image descriptor.In view of that it will cost too much time when learning quantizer or quantizing a vector in original vector space, in this paper, an approximate nearest neighbor search method is presented which is based on projected residual vector quantization (PRVQ) whose quantizer is composed of multi-stage sub-quantizer. In the process of generating sub-quantizer, the original high-dimensional vectors are projected into low-dimensional vector space by use of principal component analysis, sub-quantizer codebook is generated by kmeans clustering algorithm, and then the quantization errors are reprojected into original vector space which are used to generate next stage sub-quantizer. The above process is repeated several times to generate multi-stage sub-quantizers and the quantization errors generated in the last stage are discarded. The generated quantizer is used to encoding the image vector representation which meets well the storage constraints for large-scale image database. When a query is submitted, asymmetric distance calculation and lookup table are used to obtain the high retrieval accuracy and efficieny. In addition, by considering the overall error brought in the process of building quantizer and encoding dataset vectors, several optimization meamures are put forward to reduce quantization errors and further improve the search accuracy.In order to improve search efficiency, in this paper, we present a non-exhaustive search strategy based on two level adjacent graphs to meet large scale image retrieval. The first level adjacent graph is used to generate the query’s neighbor seeds rapidly, which reflects neighbor relationship between transformed product quantizer’s codebook and database vectors. In the process of generating the first level adjecent graph, a reverse generation mode with lower time complexity is used, in which every database vector is firstly allocated to some nearest neighbor codewords and then several database vectors with nearest neighbor relationship for every codeword are retained. The second level adjacent graph is used for neighborhood propagation, which reflects the neighborhood relationship between database vectors and themselves. In the process of generating the second level adjecent graph, a fast nearest neighbor algorithm is used to generate approximate adjacent graph of database vectors. For a given query, the nearest codeword is obtained fast, and then the first level adjacent graph is used to get the codeword’s nearest seed vectors which are regarded as the search result, meanwhile, the second level adjacent graph is used to carry out neighborhood propagation. The above process is repeated until the number of the search result equals to the specified threshold value. The non-exhaustive search based on the two levels adjacent graph provides a new index structure which solves the efficiency problem for large-scale search.There are still many problems to be studied in content-based large scale image retrieval. For the image descriptor, how to fuse various visual features to generate image descriptor needs to further study, so that the content in image is represented better. In term of fast approximate nearest neighbor retrieval, hashing is fast but with low precision. How to combine sematic information with hash to improve retrieval precision is another problem need to be studied.