Cross-media Retrieval Technology Based On Feature Association Analysis

With the wide application of information technology,the Internet has integrated into all fields of people's lives,and also brought massive growth in multimedia data.In reality,not only the types and structures of multimedia data are complex and diverse,but the expression of heterogeneous data with similar information is also diverse and inconsistent.Therefore,how to effectively mine the complex relationships of heterogeneous data in massive multimedia data has become one of the difficulties in cross-media retrieval.Cross-media retrieval refers to data retrieval between different media types.How to deal with the heterogeneous gap and semantic gap of multimedia data is the important research work in this paper.The methods we proposed are mainly carried out in deep research from feature correlation and semantic discrimination of data,fully explored the complex relationships between heterogeneous data,and accurately analyzed more comprehensive semantic information.The research work in this paper mainly includes the following two aspects:The first work of this paper is to propose a cross-media retrieval algorithm based on graph regularization and modality-dependent.Based on discriminative subspace learning,the algorithm uses the data features of different media and the corresponding semantic categories to construct a semantic association graph.Leading graph regularization aims to maintain the underlying manifold structure of different spaces,so that the feature distribution of different data between in the semantic space and in the original space tend to be consistent.In addition,the one-to-one correspondence between similar data is maintained in the mapping process,and the query objects of different media are linearly mapped in different semantic spaces,thereby learning to obtain their mapping matrix.The second research work is to propose a cross-media retrieval algorithm based on feature association learning.The algorithm not only considers the potential correlation and semantic consistency of different data,but also analyzes the potential associations between intra-media data and inter-media data by constructing isomorphic association graphs and heterogeneous association graphs of multimedia data.When multimedia data are mapped to the semantic subspace,the semantic consistency between similar data is maintained,and the feature association between different data is also maintained.In short,the above two cross-media retrieval algorithms proposed in this paper mainly aim at the spatial heterogeneity of different data,and realize the linear mapping of heterogeneous data in the common subspace.At the same time,the two algorithms make full use of semantic information and data features to construct feature association graphs of different data.A large number of simulation experiments are carried out,which verifies the feasibility and effectiveness of the two algorithms.