Research On Cross-modal Hash Retrieval Based On Common Subspace Learning

With the development of the Internet and the smartphones,a large number of multimedia data such as short videos,voices,pictures,and texts have blossomed.More and more efforts have been devoted to how to achieve efficient and accurate cross-modal data retrieval.Existing studies show that the data of different modalities can be converted into the same type of hash code by the cross-modal hashing method.It not only develops cross-modal data retrieval but also reduces the cost of storage space and time greatly.Most of the existing cross-modal hashing methods map the feature data of heterogeneous modalities into a common subspace,and then convert the real value subspace into discrete hash code through symbolic function.However such methods still have some shortcomings.For example,some methods only embed the category information of the labels into the hash codes by projection,but ignore the semantic similarity and latent semantic correlation.In addition,some methods use labels to construct large-scale similarity matrix,which greatly increases the cost of time and space.Finally,the hash code learning strategies in many methods do not depend on original features,which leads to the fact that hash codes cannot maintain the correlation between heterogeneous modalities.Based on these,we propose a similarity preserving discriminative cross-modal hashing method and an asymmetric discriminative discrete cross-modal hashing method.The main research contents are as follows.(1)The similarity preserving discriminative cross-modal hashing method fully mines the discriminative information and latent semantic correlations of the labels by the multi-label kernel discriminant analysis,and embeds them into the common subspace.In addition,the semantic similarity is obtained by calculating the inner product of labels,and then we can preserve the semantic similarity into the hash code.On this basis,we can not only improve the discriminative ability of hash codes,but also avoid the use of large-scale similarity matrix.At the same time,by minimizing the quantization error between the common subspace and the hash code,the discriminative information in the common subspace can be retained in the hash code.Finally,orthogonal and balanced constraints are imposed on the common subspace,which can further achieve better performance.(2)The asymmetric discriminative discrete cross-modal hashing method imposes discrete constraint on the common subspace,which further reduce the loss of discriminative information and quantization error.Furthermore,by incorporating latent factor models,not only discrete hash codes can be directly generated,but also semantic similarity can be preserved in hamming space.In this thesis,the two methods have conducted extensive experiments on the three most widely used datasets,and compared with the latest and most representative cross-modal hashing methods.The results show that methods proposed in this thesis have higher retrieval accuracy to other comparison methods.