Research On Semi-supervised Cross-media Feature Mapping Methods With Improved Loss Function And Joint Graph Regularization

With the development of internet and information processing technology,multimedia data increased greatly.How to transform heterogeneous multimedia data into isomorphic data-features and measure the similarity between media data with different types is the main challenge for cross-media retrieval.We discuss the improved methods of loss function and joint graph regularization and design the improved methods for semi-supervised cross-media feature mapping in this thesis.Most of the existing cross-media retrieval methods usually model similarity or semantic information,separately.Modeling with both similarity and semantic information can improve the accuracy of cross-media retrieval.The existing methods do not make full use of unlabeled multimedia data and be sensitive to noise and outliers.We optimize the performance of cross-media retrieval by improving loss function and joint regularization methods.(1)The feature mapping method is optimized by incorporating unlabeled information of multimedia data into loss function.The semantic information of unlabeled data in the optimization process is represented by feature distance.The mapping matrix of cross-media retrieval with higher recognition can be obtained by using the improved feature mapping,semantic information and joint graph regularization.(2)The reconstruction regularization constraints with l1-norm are used to improve the performance of cross-media retrieval.The sparse mapping matrices of media data with different types are studied based on common subspace theory.The loss function is regularized by l2,1-norm instead of traditional F-norm,which makes the learned mapping matrices more sparse.(3)We regularize the initial structure information of multimedia data with l1-norm graph Laplacian operator instead of the conventional graph Laplacian operator to make the model more robust to noises and outliers.On this basis,an optimization algorithm based on alternating direction method of multipliers is proposed to optimize the graph Laplacian regularization method with l1-norm.We tested the proposed methods on public datasets Wikipedia and XMedia,and compared it with existing methods.Extensive experiments show that the performance of cross-media retrieval is improved to a certain extent.