Research On Hypergraph Based Image Retrieval And Tagging Technology

With the development of social network and mobile multimedia,massive amounts of images textually annotated by different users are provided by social image websites,e.g.,Flickr.Social images are always associated with a variety of information,such as visual features,tags,and users.In general,we only managed each single feature or utilize ordinary graph for semantic network representation.However,a single feature can only measure a certain aspect of relevance rather than the true semantic association.The ordinary graph cannot represent high-order relations,resulting in information missing.Therefore,we need a framework that integrates multiple features and represents high-order relations to perform efficient retrieval and management.In this paper,instead of ordinary graph,we utilize hypergraph to model social images,since relations among various information are more sophisticated than pairwise.Based on the hypergraph,we propose HIRT,a scalable interactive image retrieval and tagging system,where Personalized PageRank is employed to measure the node similarity,and the top-k query is used to support multiple functions including the image similarity retrieval,the keyword-based image retrieval,and the image tagging.To achieve the scalability and efficiency of the system,using bulkload,paralleling,and buffering techniques,we present four efficient methods to compute the transition probability matrix that is stored by using a disk-based B+-tree,and develop paralleled and approximate personalized pagerank algorithms to accelerate top-k search.Moreover,an interactive method using the crowdsourcing technique is also presented to further improve the quality of the top-k query.Experimental analysis on a large Flickr dataset confirms the effectiveness and efficiency of our proposed system HIRT,compared with the existing system and techniques.