Research On Hash Methods For Large Scale Cross-modal Retrieval

Recently,with the rapid development of information technology,the data accumulated in all walks of life are showing an explosive growth trend,which means that we have entered the big data era.Big data is mainly reflected in two aspects:the increasing amount of data and data dimension,and the increasing number of data types.For example.Big data have a broad application in many area,and it has become an important national strategic resource.How to store,manage and analyze the big data has become a hot spot in academia and industry,big data retrieval has also become a major problem.To make effective use of big data,machine learning technology is essential.Due to the increasing amount of data and data dimension,e.g.WeChat and QQ have 900 million active users,Taobao's double eleven turnover was 168 billion,Precise Nearest Neighbor(PNN)search is hard to achieve in big data retrieve,thus Approximate Nearest Neighbor(ANN)search plays a fundamental role in big data retrieval.ANN search returns the most similar result to the input data by comparing the similarity between data.This method maintains a relatively accurate retrieval accuracy at the same time costs low time complexity.ANN search act a pivotal part in many application,e.g.cross-modal retrieval.With the increasing number of data types,in many scenarios,multiple modalities,e.g.images,texts,exist for a same object.Sina Weibo has 3.6billion pictures with text describe.The purpose of cross-modal retrieval is to find out the relationship among different modalities,then use one modality to retrieve another modality which is semantic similar to former.Cross-modal retrieval not only needs to pay attention to the correlation among modalities,but also to ensure the correlation within the modality.As the interests on searching cross-modal data increases,cross-modal retrieval becomes an emergent issue.Hashing based ANN search has attracted much attention in big data search task because it takes low storage cost and queries fast.Hashing method maps original features into a low-dimensional binary space,i.e.hamming space,and tries to maximize data information,e.g.local structure,semantic information.The similarity of data could be got quickly by computing hamming distance of hash code.Hashing method can not only reduce data storage and communication overhead,but also reduce the data dimension,thus improve retrieve efficiency significantly.Moreover,some cross-modal hashing methods have also been proposed to perform efficient search of different modalities.Some methods relax the constraint(e.g.approximation substitute)in optimizing objective function or get hash function and hash code in two step that makes information loss.Some methods using complex objective function which makes slow in training.However,there are still some problems to be further considered.For example,some of them cannot make use of label information,which contains helpful information to generate hash codes;some of them firstly relax binary constraints during optimization,then threshold continuous outputs to binary,which could generate large quantization error.To consider these problems,in this thesis,we present a supervised cross-modal hashing without relaxation(SCMH-WR).It can not only make use of label information,but also generate the final binary codes directly,i.e.,without relaxing binary constraints.Specifically,it maps different modalities into a common low-dimension subspace with preserving the similarity of labels;at the same time,it learns a rotation matrix to minimize the quantization error and gets the final binary codes.In addition,an iterative algorithm is proposed to tackle the optimization problem.SCMH-WR is tested on three benchmark data sets.Experimental results demonstrate that SCMH-WR outperforms state-of-the-art hashing methods for cross-modal search task.