Research Of Hashing-based Method For Large-scale Image Retrieval

With the rapid development of the multimedia technology,internet technology and computer hardware technology,the number of image people can get access to increases with exploration.The applications based on image provide great convenience for life and production.As an information carrier,image can deliver more information,so it is a more natural way to use it for communication.To mine the content of images and make them into well management,the first step is to retrieve the vast amounts of images.It's a challenge work to retrieve on such vast amounts of images,since the number of images is large,and the feature dimension is high,and the search time need to be reduced.The brute-force search can get good accuracy,but it's extremely time-consuming and low memory usage efficiency,so tree-based method has been proposed.The tree-based method shows promising performance,however,the performance of tree indexing drastically degrades to exhaustive linear search for high-dimensional data.Besides,it also suffers from memory constraint.In order to alleviate this problem,hashing-based method is exploited as a solution to the Approximate Nearest Neighbor(ANN)search.Compared with the tree-based method,hashing-based method maps feature into compact binary codes.Since relatively few bits are required,hashing-based method greatly reduces the storage consumption,which makes the storage efficient.In addition,hamming distance between two binary codes can be computed with milliseconds by conducting bit XOR operation,which makes the retrieval finish in a sub-linear or even constant time.This paper focuses on hashing-based method for large-scale image retrieval,and the main work and contributions can be summarized as follows:(1).This paper makes research on the typical ANN methods for image retrieval,and it summarizes their advantages and shortcomings.The existing hashing-based methods have been surveyed and classified into categories.Besides,for popular hashing-based methods,we summarize theirs main ideas.(2).A hashing method based on sparse reconstruction for image retrieval is proposed.For the reconstructed coefficient of the proposed method,a constraint of l21 norm is imposed for feature selection,so as to increase the discrimination of feature before encoding binary codes.Moreover,the sparse reconstruction is fully related with the encoding process in the framework,so that it can interact and influence with each other.Finally,a reasonable balance mechanism is built between adjusted covariance matrix and the minimization reconstruction error.As a further expansion,local neighborhood relation is constrained to the sparse reconstruction using graph Laplacian matrix,which makes sure the local structure relationship in the Hamming space preserves the local structure relationship in the Euclidean space.Therefore,it ensures the local relationship between original feature space and binary codes is the same.Compared with the mainstream popular unsupervised hashing methods,the proposed hashing method tested on several public image datasets shows the performance is improved under different evaluations.(3).A latent semantic minimal hashing for image retrieval is proposed.The motivations of the proposed hashing method has two points.Firstly,since the binary encoding process for data is a discrete domain to a continuous domain,the quantization loss should be as small as possible.Secondly,from the view of feature representation level,the existing Euclidean distance similarity hashing methods directly or indirectly reduce dimensional feature from the feature space to quantified hamming space,which lacks of feature re-refining.The feature may not represent good when it quantifys to binary codes.Taking these two considerations,the proposed hashing method using matrix decomposition to map the feature to the latent semantic feature space under the premise of minimal quantization loss,and a latent semantic minimal hashing method model is proposed in this paper.