Research On Supervised Hashing Methods For Image Retrieval And Classification

Recently,with the rapid development of society,the Internet has been more and more important in our daily life.Besides,with the popularity of electronic devices and social media,data(e.g.,texts,images and videos)is growing rapidly.Images are more popular because the strong expression power and show a trend of exponential growth.It brings great challenge to image retrieval and storage.Nearest neighbor search is the classical algorithm in information retrieval.However,it is very difficult to search the nearest neighbor precisely when the data is large.To solve the problem,approximate nearest neighbor search attracts much attention.It has been very popular in academic community because the efficiency and satisfying performance.For the similarity search task,traditional methods can work well by comparing original features when the data size is small.However,they become infeasible on large scale data because of their big computational cost of similarity calculation.Thus,in these years,a lot of efforts have been devoted to this problem.Among those proposed methods,hashing based methods have shown their great power to handle such large scale applications.Hashing based similarity search methods transform original visual features into a low dimensional binary space,while at the same time preserve the certain data properties(e.g.,local structure,semantic similarity)as much as possible.By computing the pairwise Hamming distance of binary codes,approximate nearest neighbor search can be efficiently performed on massive data corpus.They have efficient constant query time and can also reduce storage by storing compact codes of data points.Thus,in these years,hashing based methods have attracted more attention.Now many hashing methods have been proposed and some of them have good performance.However,most of them are proposed for retrieval task instead of classification.This means that we cannot predict the labels with the obtained binary codes.It's a kind of waste.Thus,if a hashing method can efficiently tackle retrieval and classification tasks simultaneously,it would be much useful and practical.Motivated by this,a novel efficient supervised hashing method,i.e.,Class Graph Preserving Hashing(CGPH)is proposed,which can well embed semantic information into hash functions and directly predict the labels of query data using the learned hashing codes.Specifically,CGPH first incorporates the supervised label information for jointly learning effective hash functions and correlations between tags and hashing codes to preserve the consistency between labels and hashing codes.In addition,to preserve the class similarity between samples,it also construct the class mapping based on manifold regularization by semantic neighbor graph.Experimental results on three real-world image data sets with semantic labels show that CGPH can outperform or is comparable to other state-of-the-art methods for image retrieval and classification tasks.However,labeling samples requires much human expertise on large scale data set.For this reason,available supervised information can be very limited.Semi-supervised methods will be helpful in this case.They often take advantage of both supervised information and underlying similarity information of data.Nowadays,to optimize the objective functions more easily,the semi-supervised hashing methods often relax the discrete hashing codes first.It causes some information loss.Moreover,to use the image structure fully,some methods use the similarity matrix to preserve the similarity,which is nxn.It's hard to compute and store,even can't run on some large data sets.Motivated by this,a semi-supervised hashing with graph cut(SHGC)is proposed,which optimizes the discrete hashing codes with graph cut and embeds the sparse expression into the data to reduce the dimension.The experiments are conducted on two data sets.It shows that SHGC outperforms other methods in most cases.