Research On Cross-modal Hashing Retrieval Method Based On Co-training

With the rapid development of multimedia devices,Internet,and cloud computing,the information society has entered the era of big data.The multimedia data including text,image,audio,video,etc.,has various sources and modalities.As the carrier of information transmission,the multimedia data contains abundant economic and social value.In order to make the best of the information obtained from the multimedia data,how to efficiently analyze,manage and store these data,has become a highly concerned problem for researchers and engineers.As one of the famous retrieval methods,hashing technology is widely used as its remarkable efficiency gains and storage reductions.Thus,cross-modal retrieval method based on hashing has great research significance.In order to match the different features in cross-modal data,hashing based cross-modal retrieval methods pay more attention to project the original data from different modalities into a common low-dimensional subspace.Then it is convenient to measure the similarity of different modalities in the low-dimensional subspace.However,most of them do not make full use of the discriminative information between the different modalities,thus they can not catch the the inherent discriminative structure of the data.With the help of some related knowledge such as matrix factorization,subspace learning,co-training,and graph embedding,promotions and improvements are proposed in this thesis to solve the above problem.The main research contributions are summarized as follows.Firstly,a locally linear cross-modal hashing method based on sparse subspace learning is proposed.Considering collective matrix factorization hashing ignoring intra-similarity of the cross-modal data,we propose a locally linear cross-modal hashing method to complete the original framework.To this end,sparse subspace learning is used to learn the sparse low-dimensional representation,which regarded as the intra-similarity constraint.The unified hash code is obtained by quantizing low-dimensional representation coefficient,which generated in the process of matrix factorization.Secondly,a co-training based matrix factorization cross-modal hashing method is proposed for taking full use of effective inherent discriminative structure information contained in the cross-modal data.In this framework,collective matrix factorization is utilized to project the data from different modalities into a common subspace,and the low-dimensional latent semantic representation of original data can be obtained.O n this basis,inter-modal similarity can be preserved via co-training method,which constrains the similarity of one modality through the discriminative information obtained from the other modality.And the intra-modal similarity can be preserved by k-nearest neighbor method simultaneously.Then,the discriminative hash code can be obtained by quantizing low-dimensional representation coefficient.Experiment results show that compared with the existing unsupervised cross-modal hashing method,the two methods based on sparse subspace learning and co-training which proposed in this thesis can improve retrieval accuracy and show the superior retrieval performance.