Research Of Re-ranking Method For Multimedia Retrieval

Recent years,re-ranking methods have attracted huge attention for their great power to improve accuracy of multimedia retrieval.Among all kinds of re-ranking methods,graph-based methods have been being mainstream.Graph-based re-ranking methods can roughly be divided into two categories: global context-sensitive(GCS)method and local context-sensitive(LCS)method.With the explosive growth of multimedia data,LCS re-ranking methods are becoming popular for their lower time complexity.LCS methods commonly have two steps: 1)remeasure similarities between multimedia objects utilizing LCS similarity measurement;2)select multimedia objects as new rank result according to new similarity relation.In this paper,two methods corresponding to these two steps are proposed.At the stage of remeasuring similarities,traditional methods always take one layer of neighbors as context information,which are sensitive to the size of neighborhood.If the size of neighborhood is too small,similarities of some multimedia objects can't be measured.And,if the size of neighborhood is too large,much noise may be brought in.To solve this problem,a new similarity measurement named LSNRI is proposed in this paper.In this new method,multiple layers of neighbors and ranking information of original ranking list will be used when measuring similarities.By the means,not only more context information can be exploited but also less noise will be introduced.Another method proposed in this paper is a cluster-based re-ranking method named Greedy Spectral Clustering(GSC),which serves in re-ranking stage.GSC is based on spectral clustering method but more efficient.It is a greed solution of spectral clustering motivated by the specialness of multimedia retrieval problem.To compensate the weakness of GSC algorithm that GSC is sensitive to the first few selected objects,query expansion is used to initialize GSC algorithm.Conbining GSC and QE,our method can improve accuracy of multimedia retrieval further on the base of LSNRI.To validate methods proposed in this paper,experiments conducted on five image and music datasets are presented,including Corel-1K,Corel-10 K,Coil-100,UKbench and GTZAN.Experimental results show that our method can improve accuracy of multimedia retrieval effectively.