Research On Cross-media Retrieval Technology Based On Dictionary Learning

With the rapid growth of multimedia data on the internet,the research on the management of multimedia data is now a hot field.Traditional multimedia research mainly focuses on multimedia data of single modality,such as image retrieval,text retrieval,audio retrieval or video retrieval,and cross-media data representation learning is mostly ignored.Although huge amount of independent visual and textual data are available today,only a small portion of them are linked with semantic associations.Motivated by big data and diversity of data forms,multimedia data with massive volumes and high dimensions are pervasive,and data sharing is becoming more and more urgent,more and more researchers focus on analyzing the correlation of data among different modalities.In this situation,the cross modal retrieval emerged.The similarity measurement of different types of media objects has always been a challenge.One of the most important problems is the heterogeneity gap across domains that original features from different modalities have different physical meanings and dimensions.There also exists a semantic gap between the low-level original features of multimedia content and high-level semantic when the user retrieves,which is a difficult problem in computer similarity between heterogeneous media objects.To solve these problems,two kind of cross-meida retrieval algorithms based on subspace learning are proposed,Experiments on multiple benchmark datasets show the effectiveness of our approach.The main work is summarized in two aspects as follows:(1)we propose a novel approach to cross-media retrieval framework based on Linear Discriminant Analysis(Cross-media retrieval based on linear discriminant analysis).This method satrts from subspace learning,which integrates the correlation between textual features and visual features.Thus the discriminative characteristic of textual modality is transfferred to the corresponding visual features via the correlation analysis process.(2)we propose a cross-modal retrieval tasks based on dictionary learning with common label alignment(Cross-media Retrieval Technology Based on Dictionary Learning).Based on the subspace learning,the method conducts coupled dictionary learning on the data from different modal and then projects them into a common space.where the correlation between these modalities is encouraged by using common label alignment.Extensive experiments show that our proposed method yields state-of-the-art result on three benchmark datasets.