Research On Shape Matching Methods Based On The Relationship Of Contour Points Orders

With the rapid development of computer technology, the research of shape matching methods has become a hotspot in the field of computer vision and pattern recognition. And many methods have been successfully used in face recognition, robot navigation, medical image analysis and other fields. The matching of shapes means that using reasonable metrics to calculate the distance or similarity. So the main focus of shape matching are the construction of shape descriptor and selection of feature matching method.This paper analyzes the pairwise matching methods and shape distance learning methods. In the pairwise matching methods, in order to improve the robustness and discrimination power of the triangle-area representation, a shape matching method based on multi-scaled contour space relationship is proposed. The multi-scaled information can be obtained by analyzing the spatial relationship of different points, and a new shape descriptor which can express the local and global information of shapes is further constructed. As the orders of all the contour points are already known, dynamic programming method and shape complexity can be used to calculate shape matching results during the feature matching step. Besides, shape distance learning is employed to improve the precision of pairwise matching methods, in which the manifold structure of the samples is not considered. The mean first-passage time (MFPT) can update the distance obtained by pairwise method and effectively improve the retrieval rates. However, when the distance of two samples which are in the same category is large, the mean transition time will be large, and the updated distance could be inaccuracy. Thus a novel shape distance learning method based on generalized mean first-passage time (GMFPT) is proposed to solve this problem. Given a set of shapes as the state space, generalized mean first-passage time is used to represent the average time steps from one state to a certain set of states, and the shortest paths on the distance manifold can be explicitly captured. Numerical experiments on different datasets demonstrate that methods proposed in this paper can achieve desirable results effectively.