Research On Data Method Of Cross-modal Discrete Hash Learning

With massive and diversified data flooding all around world today,how to retrieve the target data from the complex and huge data has become an important issue.It needs to be solve urgently in the retrieval direction.Hash retrieval has attracted widespread research attention due to its advantages of accuracy and speed in retrieval.Studies have proved that supervised hashing methods are more effective than unsupervised hashing methods in actual retrieval applications.Although supervised cross-modal hashing technology has made considerable progress,there are still some problems need to be solved.For example,in order to obtain the modal similarity,most supervised hashing methods use the paired similarity matrix,which will increase the computational complexity of the method and consume huge storage costs.There are some hashing methods how to use different modalities.The similarity of modal forces the modal information be mapped into the same latent semantic space and neglects the original information contained in the modal,resulting in the lack of semantic information.In addition,some methods relax the hash code binary constraint during model learning.This also has the problem of high retrieval cost and low accuracy.The main research contents of this paper are as follows:(1)Use the two-step hash method to learn the hash code and hash function step by step,making the entire hash learning process more flexible.In order to be able to learn a richer common potential subspace and reduce the retrieval speed,we use Distance-Distance Difference Minimization Problem to embed the rich semantic features between the original modal information and the label information into the subspace.By avoiding the use of large-scale paired similarity matrices,time and computational complexity are reduced,and the accuracy of the search can be improved while also improving the retrieval speed.(2)A valid hash code can be generated even when the constraint is relaxed.Since the generated hash code is discrete,it becomes difficult to solve it after imposing some conditional constraints on it.We impose constraints on the latent subspace to make it approximate to B.(3)Use multiple subspaces to learn the similarity and unique information of different modalities.In order to ensure the similarity between different modal data,while maintaining the unique information of different modalities,we use multiple subspaces to learn different The modal information keeps the structural characteristics of the original modal in the multiple subspaces learned.The experimental comparison further verifies the effectiveness and accuracy of the method.The two methods proposed in this paper are compared with the current manifold benchmark methods in cross-modal hash retrieval.Experiments show that the performance of the proposed method is better than the current benchmark methods.