Research On Cross-media Retrieval Algorithm Based On Discriminative Subspace Learning

At present,multimedia data is growing explosively.Different types of data usually appear in different modalities in Web pages,describing the same topic or event from different perspectives.How to abstract and associate different modalities with the same semantic information has received extensive attention from the academic and industrial areas.In practical applications,people usually expect to retrieve related results from other modalities by submitting query data of arbitrary modalities,and achieve the retrieval between different modalities.In order to meet such practical demand,cross-media retrieval technology emerges as the times require.Since the data of different modalities usually exist in heterogeneous feature spaces,the semantic "gap" appears.How to bridge the semantic "gap" between the underlying features and the high-level concepts of multimedia data,and evaluate the similarity between multimedia data is a major challenge of cross-media retrieval technology.Based on the subspace learning,this paper further explores the correlation between multi-modal data,and learns more discriminative common subspace to improve cross-media retrieval performance.This paper proposes three cross-media retrieval algorithms.The main work and innovations are as follows:1.This paper proposes a cross-media retrieval algorithm based on joint feature selection and graph regularization.When learning common subspace,the algorithm considers the differences between different sub-retrieval tasks and the discriminant information embedded in multi-modal features.It seamlessly integrates linear regression item,correlation analysis item,graph regularization item and feature selection item into a joint cross-modal retrieval framework.These different items interact with each other and embed rich semantic information into the common subspace.Experimental analysis on several public datasets shows that the algorithm achieves superior performance.2.This paper proposes a task-dependent and query-dependent cross-media retrieval algorithm.In practical applications,query data has semantic diversity,which may come from categories with unique semantic distributions.To solve this problem,in the offline phase,the algorithm learns projection matrices for specific tasks and categories through joint optimization,and trains effective classifiers to obtain semantic mapping functions.In the online retrieval phase,the optimal projection matrices are adaptively determined by combining the specific retrieval tasks,the potential semantic categories of query data and the task-category-projection mapping table.The multi-modal data is mapped to the common subspace for similarity measurement and the retrieval results are returned.3.This paper proposes a fusion-supervised deep cross-media hashing algorithm.This algorithm combines hash technology and deep learning to learn unified hash codes through the fusion hash network,which effectively enhances the non-linear correlation between multi-modal data.On this basis,the unified hash codes are used as supervised information totrain the modality-specific hash networks,and the pairwise similarity and classification information are embedded into the hash networks.Thus,cross-modal similarity and semantic consistency can be preserved more effectively.The experimental results on two public datasets show the superiority of this algorithm.