Cross-modal Hashing Retrieval Based On Hierarchical Structure Modeling

The purpose of cross-modal retrieval is to find the correlation between data from different modalities.Hashing learning,as a kind of popular methods with high efficiency,high effectiveness and low storage cost,have been increasingly applied to cross-modal retrieval domain.However,most cross-modal hashing methods try to learn the hash functions by mapping the original features of different modalities to a common hash coding space,and relaxing the binary constraints on their own hash spaces,which cause the representation of each modal is not accurate enough.In order to reduce the computational costs,some algorithms used randomly selected data from entire training set to learn the hash functions,which result in randomness of query results and also missing some important samples.Aiming at the above problems,based on the cross-modal hashing retrieval methods,this thesis proposes adaptive cross-modal fast hashing method,density-based cross-modal variable-length hashing,and cross-modal variable-length hashing based on hierarchy,respectively.The effectiveness of these algorithms is verified from the theoretical and experimental aspects.The specific researches are as follows:1.In view of the problem that traditional cross-modal hashing methods have high costs of training process,and query performances are not accurate enough,this thesis proposes a Cross-modal Hashing Based on Density Clustering(DCCH)method.DCCH combines density clustering algorithm to adaptively select representative images and texts,then choose the representative samples both satisfying two modalities,and respectively project them to the common hash coding space.Finally obtain hash functions of each modal by discrete cross-modal hashing methods and generate hash codes.Experimental results demonstrate this method can effectively improve the accuracy of cross-modal retrieval,while greatly reducing the training time.2.Since the traditional cross-modal supervised hashing algorithms do not pay attention to semantic similarity within modal and does not change the length of hash codes according different modalities,this thesis improves DCCH and proposes a Cross-modal Hashing Based on Density Peak Clustering(DPCCH).DPCCH selects representative image-text pairs through the overall density distribution,and then respectively projects to different length hashing codes to better represent different modalities,and then iteratively generates hashing codes of each modal.Experimental results prove that DPCCH further improves the retrieval performance of cross-modal data.3.In order to solve the traditional cross-modal supervised hashing methods without considering the similarity of original data in the manifold structure,resulting in the problem of excessive computational costs and low retrieval performance,this thesis proposes a Cross-modal Variable-length Hashing based on Hierarchy(CVHH).CVHH combines manifold learning and hierarchical theory to construct the hierarchical structure and only use top-layer representative samples to learning hash functions and generating hash codes.Experimental results illustrate CVHH can effectively improve the retreival performance and also reduce the training time.