Research Of Re-ranking Method For Content-based Multimedia Retrieval

With the rapid growth of multimedia information,people are increasingly demanding effi-ciency and precision of multimedia retrieval.In the process of multimedia retrieval,the length of features determines the efficiency of most retrieval methods.Therefore,how to effectively reduce the feature dimension without losing similarity relationship between features has become one of the problems that researchers need to solve urgently.At the same time,the graph-based multimedia retrieval re-rank method has also received extensive attention because it can effec-tively reduce the semantic gap in the retrieval process.Generally speaking,graph-based re-rank methods are divided into two types: global-based re-rank methods and local-based re-rank meth-ods.This paper proposes a method for the dimensionality reduction and a local-based re-rank method.Multimedia features are always distributed on manifolds in high-dimensional space.Lin-ear dimensionality reduction methods often cannot retain the similarity between multimedia ob-jects.In contrast,nonlinear dimensionality reduction is usually a better choice.Local linear embedding(LLE)retains the local nearest neighbor linear relationship in the original space in the low-dimensional space,and often achieves better results in the graph-based multimedia re-rank method.However,the LLE is not suitable for the retrieval environment because it cannot achieve the out-of-sample dimensionality reduction.In order to solve this problem,this paper proposes an out-of-sample LLE dimensionality reduction method based on RVFL.On the other hand,this paper proposes a similarity measure method named local neigh-bors ranking information(LNRI).The method is inspired by the neighbor relationship in social networks,and utilizes the ranking information of local neighbors to re-measure the similarity be-tween multimedia objects.Through the neighbor ranking information,the algorithm effectively utilizes the context information of the multimedia objects,thereby narrowing the semantic gap in the retrieval process.To validate methods proposed in this paper,experiments conducted on seven multimedia datasets are presented.Experimental results show that the method proposed can effectively im-prove retrieval efficiency and accuracy.