Research On Shape Retrieval Methods Based On Self-adaptive Fusion For Distance Learning

With the rapid development of society,computer vision has played an increasingly important role in many aspects of human production and life.In the research of computer vision,shape matching and shape retrieval have always been problems people focus.Solving such kind of problem can improve the application of computer vision in different fields.Shape retrieval methods usually include shape feature extraction,shape feature matching and shape distance learning.The first two steps constitute pairwise shape matching method,and the third step is the process of re-ranking on the basis of the first two.All the above three steps have a direct impact on the shape retrieval result.Many scholars have done a lot of work about shape retrieval in different perspectives.This paper mainly studies shape feature matching step and shape distance learning.For global shape contour features,the matching results are often calculated by analyzing the correspondences between different contour segments.It has always been a concern to combine the shape features characteristics and analyze the corresponding relationship between fragments effectively.Taking shape feature as object based on contour point space location,we consider introducing spectral clustering algorithm to analyze the correspondence between different contour sampling points.On the basis of shape feature extraction,we analyzed the distance relationships between shape descriptors corresponding to sampling points of different contours.Further,we use spectral clustering method to check the correct correspondence between the sampling points in different contours and delete redundant connections.Experiments on different datasets show that the method is effective.As it is difficult to capture the context information effectively in shape dataset by shape matching method,many researchers have introduced distance learning into shape retrieval as a post-processing step.In recent years,considering the differences in emphasis focused by different shape features,more and more studies use the fusion for distance learning method as the post processing step.Based on this,the results of matching shape matching are re-ranked.Considering the different roles played in each feature,this paper proposes a self-adaptive fusion for distance learning method to solve this problem.In the process of re-ranking,we analyze the relationship between different similarity spaces,then use a semi-supervised framework to analyze the context information in different similarity spaces.Experiments on different datasets verify that the self-adaptive fusion for distance learning method is effective.