Research On Cross-modal Retrieval Based On Semantic Discriminative Hash

With the development of internet technology and the rapid flow of information,multimedia data has grown rapidly,showing the characteristics of larger scale and increased types.In this case,it is difficult for traditional uni-modal retrieval to satisfy users' retrieval needs.At the same time,large-scale data bring new challenges for cross-modal retrieval.Utilizing the characteristics of low storage and high efficiency retrieval of hash codes,crossmodal hash retrieval has attracted more and more attention from cross-modal retrieval researchers.The core issue of cross-modal hash retrieval is how to use potential correlations in heterogeneous data to shorten the semantic gap.Most methods ignore the interactive exploration of potential semantic related information,and only use matrix which contains binary values to indicate the degree of correlation,unable to capture deeper semantic information between multi-label data,and ignore the maintenance of semantic structure and the discriminative characteristics of data,thus affecting the performance of cross-modal retrieval.In order to fully mine the potential semantic related information between heterogeneous data and learn discriminative hash codes,the cross-modal hash retrieval method based on deep discriminative semantic union proposed in this paper uses the common information between the data and their unique complementary information to mine the deeper semantic associations of the data;and then uses the classifier to learn discriminative hash coding.In addition,the proposed discriminative cross-modal hash retrieval based on multi-level semantics uses a multi-level semantic similarity matrix to measure the different degrees of correlation between cross-modal data on the basis of deep hashing,so that the learned hash code is distinguished while maintaining high-level semantic correlations.On two commonly used datasets,precision recall curve and mean accuracy precision are used as the evaluation criteria,the experimental results in this paper confirm that the method proposed in this paper improves the performance of cross-modal retrieval.