Research On Shape Matching Methods Based On Contour And Region Information

Shape is a high level visual information,in the field of computer vision research,shape is an important attribute to describe an object.Shape representation and matching as a key and basic problems in the field of computer vision which have played an important role in many fields such as object detection,medical analysis and ancient writing research.In various shape matching methods,methods based on both contour and region information have better performance because of the richness of the shape information,and the corresponding descriptors have some advantages in expression ability.In this paper,we start from the two key issues of "pairwise shape matching method" and "shape distance learning method" to study the shape matching methods based on both contour and region information.Based on shape representation methods combining contour and region information,this paper improves the aspect shape context method(ASC)and proposes a joint distance learning method based on generalized mean first-passage time(Co-GMFPT).The main research contents are as follows:(1)The traditional aspect shape context method can describe the relationship between contour sampling points,but the geodesic distance calculation process is complex and the feature extraction efficiency is low,and only select one optimal aspect space to solve the shape distance has some limitations.In view of the shortcomings of the method,this paper improves ASC method from improving the feature extraction efficiency,enhancing the descriptor expression ability and optimizing the feature matching algorithm.Proposed method transforms the problem of geodesic distance calculation between contour sample points in aspect space into the shortest path problem,and the shape descriptors can be constructed quickly by introducing shortest path algorithm,so the feature extraction efficiency can be greatly improved.Besides,fuzzy histogram is introduced to construct aspect fuzzy shape context(AFSC)to further enhance the expression ability of the shape descriptor.On this basis,shape feature matching is carried out based on dynamic programming method.According to the local features of the shape to optimize the selection of aspect space,and then analyze the more accurate shape distance relationship,and then shape distance relationship can be analyzed more accurate.The experimental results both of comparison of shape feature extraction efficiency and accuracy of the shape retrieval under different data sets show that the proposed method has good performance.(2)Pairwise shape matching methods focus on analyzing the relationship between two matching shapes,ignoring the potential data manifolds in the samples.To compensate for the shortcomings,shape distance learning algorithm is introduced to capture the potential manifold structure,and to improve the accuracy of shape matching results.Among them,co-training learning method with different distance measurement to obtain more effective information,often can get a better distance learning effect.In this paper,Co-GMFPT method is proposed by combining the advantages of fusion perspective and generalized mean first-passage time.This method considers the set of shape samples as the state space,and introduces the generalized mean first-passage time to exploit the shortest path in the sample space manifold.Related samples of the query can be found from near to far through a iterative process alternately utilizing two distance measurements.The experimental results under different data sets verify that the proposed method has good performance.