Research On Shape Matching Methods Based On Spatial Configuration Features Of Contour Points

Shape matching is one of the important issues in the fields of computer vision and pat-tern recognition for the latest half century. In various types of shape matching methods, those based on the spatial configuration features of contour points provide dynamic performances and have received widely attention. The key point of the shape matching researches is about how to improve results and efficiencies of shape matching methods by analyzing the characteristics of shape descriptors.The research work of this thesis mainly contains the following aspects:(1) As the contour information can not be accurately described by shape context, a novel shape matching method based on fuzzy shape context is proposed. Although both of the global information and the local information are obtained by building the log-polar histogram, the dis-tribution of sample points, in many cases, can not be expressed by the shape descriptors. This is because the sample points from the shape contours are directly classified into different bin-s. Therefore, the log-polar fuzzy histogram are constructed for representing the distribution of sample points, and a novel shape descriptor named fuzzy shape context is proposed. The infor-mation of shape contours can be reflected by the shape descriptor precisely, which is obtained by accumulating the fuzzy membership degrees of sample points for different fuzzy subsets. Lo-cally constraint matching method and point-set segmented matching method are proposed for analyzing the correspondences between sample points, and both of them can be used to solve the point matching problem. The circular shift matching is further designed to deal with the rota-tion invariant problem. Experimental results prove that desirable shape matching results can be effectively achieved by using the proposed methods.(2) To improve the results obtained from pairwise shape similarity analysis, measure sub-stitution and shape distance learning are employed into the shape matching process, then shape context matching method based on histogram-based earth mover’s distance and shape distance learning based on mean first-passage time are proposed. By analyzing the structure of histogram-s, a more efficient cross-bin dissimilarity measure named histogram-based earth mover’s distance is proposed by simplifying the model of earth mover’s distance. As histogram distances can not be evaluated accurately by using bin-by-bin dissimilarity measures, the histogram-based earth mover’s distance is introduced into the procedure of shape feature matching to solve this prob-lem. The experimental results show that, compared with other methods, the proposed model is less time consuming, and outstanding results are obtained by introducing the histogram-based earth mover’s distance into shape matching methods. Moreover, to avoid the imbalance of the process of label propagation, the mean first-passage time is introduced to update the shape dis-tance obtained by using pairwise shape matching method. With the Discrete-time Markov Chain obtained by analyzing the shape distance matrix, the shape distance is represented by the average time that a particle spends to finish a state transition in the state space, and the manifold of data space can be effectively captured. Experimental results of shape recognition and shape retrieval are effectively improved by using the proposed shape distance learning method.(3) To enhance the efficiency of shape matching methods, a fast shape matching method based on corner fuzzy shape context is proposed, and shape matching methods based on spatial configuration features of contour points are further introduced to solve the problem of human action recognition. With a small number of fuzzy shape contexts generated by taking the corner points as references, and a novel shape descriptor named fuzzy corner shape context is pro-posed in this thesis. The important shape information can be reflected by the proposed shape description, and the corresponding point matching time is significantly shortened. A fast shape matching method based on the corner fuzzy shape context is further presented, and the efficiency of the proposed method is demonstrated through the experimental results. Meanwhile, as human silhouettes of video images contain lots of shape information, human actions can be classified by analyzing the shapes of silhouettes. With the differences of silhouettes represented by spa-tial configuration features of contour points, shape matching methods based on the polar fuzzy histogram are used for fast pruning, and shape matching methods based on fuzzy shape context are employed for precisely matching. Numerical experiments show that the proposed method achieves comparable human action recognition results.