Research On Contour Based Shape Matching

Shape matching is to measure the similarity between two shapes with some rules. It is a fundamental problem in automatic image recognition and understanding, and is widely used in computer vision, pattern recognition, remote sensing image analyzing, handwriting recognition, etc. During the past few decades, many approaches have been proposed to tackle it. But there are still several shape matching problems unsolved, such as occlusion, non-rigid deformation and so on. This thesis studies non-rigid shape matching problems with contour representation, mainly including shape matching based on shape contexts and neighborhood, part-based shape matching, etc. The major contributions of the thesis are as follows.1.The general processes and some typical approaches of shape matching are presented. Then the advantages and disadvantages of these approaches are analyzed and compared in detail.2.A shape matching method based on local properties of object deformation is proposed, which preserves neighborhood structures in shapes. Shapes are first represented by point sets, and then the neighborhood of each point in one shape is constructed and weighted by distance. We formulate the point matching problem as an Attributed Relational Graph matching problem and solve it by relaxation labeling at last. The proposed method uses the property of non-rigid deformation that the local continuity is generally well preserved due to physical constraints. So it can improve the matching accuracy and the converging speed of relaxation labeling, which is demonstrated by experimental results.3.A shape matching method based on local search is proposed. Shape context is introduced as an attribute of point in shape. Firstly, we initial the matching with Hungarian method,and calculate the confidence for two corresponding points by their shape context distance and the support of their neighborhoods. Then we modify the correspondence of points with low confidence by heuristic search. Higher accuracy in non-rigid shape matching is achieved by combining the global characteristics of shape context and the local properties of neighborhood, which is demonstrated in the experiments presented.4.A shape matching method based on visual parts is proposed according to the human vision theory. We predigest shape contour by discrete curve evolution algorithm, and divide visual parts from it. Then we sample the contour into point set, and construct the global and local shape contexts of visual parts. With the cost matrix obtained from the two shape contexts, shapes are matched by visual parts using Hungarian method. Taking the advantage of visual parts, the presented part-based shape matching fits the mode of human vision.