Research On Cross Media Retrieval Algorithm Based On Embedded Spatial Representation

In the current world,the expression of information is becoming more and more diverse.A large amount of information,especially multi-modal information,is published on the network every day.Cross media retrieval is facing the national strategic needs of big data application and cyberspace security.The technology and application of multimedia content understanding are of great significance in the fields of harmful information identification,intelligent medical treatment,hot event analysis,multi-modal data utilization and military strategic analysis.How to manage and use the massive multi-modal data is a problem that people need to consider at present.Cross media retrieval refers to submitting data of any modal and obtaining data of different modals with similar semantics.Now cross-media retrieval is still facing many challenges.On the one hand,the underlying feature dimensions and attributes of different forms of cross-media data are quite different,so it is difficult to directly measure the similarity of underlying features between data,that is,cross-media heterogeneous differences.On the other hand,the semantic information of media data is abstract,and the semantic association between different forms of media data is abstract.In order to solve the above problems,this paper makes an in-depth research on different retrieval tasks,analyzes the semantic information and correlation between different media features,and proposes the idea of embedded spatial feature representation and modal dependency.Embedded spatial feature representation is to optimize the underlying multi-modal data features in a certain way to form an embedded spatial representation and enhance the spatial projection from the embedded text to get a more effective target matrix.In the embedded space,not only the more accurate cross media retrieval tasks can be performed,but also the fine-grained retrieval can be carried out.Modal dependence refers to learning different matrices for different retrieval tasks,and it uses similarity matrix for weighting optimization to ensure the accuracy of retrieving modal features in the regression processes so as to improve retrieval efficiency.The main contributions of this paper are as follows:(1)Text feature depth aggregation.In view of the problem that there is too much noise in the process of image feature extraction,text features often have strong discriminative ability.Considering the joint optimization of the correlation between image features and text features,the strong discrimination of text features is transferred to image features through subspace by linear discriminant analysis to further improve the retrieval efficiency.(2)Joint optimization of image features.In view of the problem of noise and semantic gap in image features,this paper combines two popular dimensionality reduction methods to optimize image features both in the whole and the local part,which effectively improves the discriminative ability of image features.