Research On Cross-media Retrieval Methods Based On Reconstruction Regular Constraints

With the rapid development of Internet technology,online resources are becoming richer and richer.The storage form of network information is more diverse,from text data to image,audio,video and 3D models.While these massive data bring people a lot of information,they also bring pressure to people to quickly query the required information.Therefore,how to achieve accurate and effective mutual retrieval between different modal data has become an urgent problem to be resolved.Because different types of data have different representations of features,their dimension in the original space is also different,they can not directly search each other.Hence,how to fully understand the semantic information of different types of media data,to establish the correlation between heterogeneous data,and to further explore the distribution information of media data in the ground space,then transform heterogeneous media data into homogeneous data for similarity measurement,which becomes the key to cross-media retrieval problems.This paper takes different retrieval tasks as the breakthrough point,and analyzes the correlation and semantic information between different media data.Two kinds of cross-modal retrieval algorithms are proposed,and experiments are carried out on different data sets to verify the effectiveness of the proposed methods.The main research work in this paper is as follows:(1)A cross-media retrieval algorithm based on query modality and semi-supervised regularization(Cross-media Retrieval based on Query Modality and Semi-supervised Regularization,QMSR)is proposed.This method starts from different retrieval tasks,and integrates the correlation between different modal data,the semantic information and distribution information of query modality in the original feature space to learn different mapping matrix,which makes the understanding of the query object more accurate.(2)A cross-media retrieval method based on modality-dependent and semi-supervised coupled dictionary learning(Modality-Dependent and Semi-supervised Coupled Dictionary Learning for Cross-media Retrieval,MDSCDL)is proposed.This algorithm integrates two processes of dictionary learning and feature mapping.In the process of dictionary learning,the distribution information of original data is converted to the corresponding sparse coefficient.In the process of feature mapping,based on different retrieval tasks,the correlation and semantic information of data are integrated with sparse representation,which makes the mapping matrix learned more pertinent.