Research On Cross-Modal Retrieval Method Based On Asymmetric Hashing

With the rapid development of mobile internet,internet of things,cloud computing,and cloud storage,the amounts of data have grown exponentially,and the information society has entered the era of big data.The enormous multimedia data coming from a wide range of sources,contains rich economic value and social value.The rapid growth of data brings new opportunities and challenges for the social development.In order to excavate information from multimedia data,how to efficiently store,process and analyze these data,has become an urgent problem in the process of big data research.Hash learning-based approximate nearest neighbor search is a popular method in the data retrieval.Hashing has been widely used as its properties of quick searching and small memory footprint.How to efficiently search these multimedia data remains to be further studied.Therefore,it is necessary to further explore hashing methods for cross-modal retrieval.Most cross-modal hashing methods focus on maintaining correlation between different modalities,and ignore the generalization capabilities of hashing methods and the complicated distribution of multimodal data.After studying the dictionary learning,non-parametric Bayesian model and asymmetric hashing in more detail,and considering the deficiencies of existing cross-modal hashing,this dissertation proposes two cross-modal hashing methods to improve the performance of cross-modal retrieval.The main contributions of this dissertation are summarized as follows.1.Considering existing cross-modal methods solving maximum inner product search hard,and easily overfitting and ignoring the insufficiency of the generalization ability,we propose an asymmetric cross-modal hashing method.In this asymmetric framework,it learns two distinct hash functions for each of the modality to efficiently solve the maximizing inner product search problem and increase the generalization ability of cross-modal retrieval.This method can effectively prevent overfitting.At the same time,this framework combines the process of encoding with the dictionary learning,which assumes that the paired data has the same or relevant representation coefficients.Ultimately,the unified hash codes are obtained and can maintain the semantic relevance of different modalities well.2.Considering existing cross-modal methods assuming the data obeys Gaussian distribution,which cannot accurately describe the distribution of multimodal data,we propose a supervised cross-modal hashing method based on non-parametric Bayesian model.In this framework,it models multimodal data using the Dirichlet process as the priori distribution of the mean of data category.At the same time,it establishes a probabilistic graphical model which can describe the association between original data,semantic labels and hash codes.Then it obtains the posterior probability between original data and the hash codes.In the end,the unified hash codes are obtained and can approximate the original data well.Extensive experimental results show that the proposed asymmetric cross-modal hashing significantly increases the generalization ability and flexibility of cross-modal retrieval.And the supervised cross-modal hashing method based on non-parametric Bayesian model describes the distribution of multimodal data well.Both of the methods improve the hash retrieval accuracy and show superior retrieval performance.