Cross-moda1 Hash Retrieval For Label Consistency Preservation

Today,with the rapid development of information technology,there are a large number of data with different modes around us,such as text,image,small video and other types of data.These multi-mode data often have higher dimensions,which makes it difficult to conduct efficient retrieval in these data.However,hashing learning converts data into binary code,which can greatly reduce the time consumption and space consumption of retrieval.Therefore,hashing learning has attracted the attention of a large number of researchers.Now hash learning has been developed for a period of time,but there are still some problems.Some methods,for example,may use subspace,manifold learning,collaborative matrix decomposition techniques such as different modal data feature extraction and save the son in a public space,but they are more focused on the extraction mode of heterogeneous data information,and less focus on the label information extraction,due to the label contains rich semantic information,The quality of hash code generated by these methods is lost.In addition,there are other methods that do not constrain the hash code when learning the hash code,or even relax the discrete constraint,making the hash code quality lower.Some studies have been carried out to address these issues:The main research contents of this thesis are as follows:(1)In order to solve some method to extract the information in the tag due to less semantic losses caused problems,our method by using the methods of manifold learning,build the similarity matrix,the tag contained in the semantic information extraction,and reuse information construction constraints,make can remain in the hash code semantic information in the label,in order to keep the label consistency,Meanwhile,label continuity is maintained to reduce semantic loss caused by constructing similarity matrix and further increase semantic information in hash code.Finally,the kernel function is used in the learning step of hash function,which improves the quality of hash function.(2)Some methods in the learning process of hash code not the constraints on the hash code,and even relax the discrete constraints,thereby reducing the quality of the hash code,in order to solve this problem,our method in the process of the hash code of optimization on the hash code constraints,and makes the generated hash code can contain more information as possible,improve the quality of the hash code.At the same time,hash learning is divided into hash code learning and hash function learning to improve the overall flexibility of hash learning and reduce the complexity of the model.(3)In order to verify the effectiveness of the proposed method,a large number of experiments are carried out on several benchmark data sets,and the experimental results of the proposed method are compared with those of several cross-modal hashing methods.The experimental results verify the effectiveness of our method.