MATCHING AND RECOGNITION OF DEFORMED CLOSED CONTOURS BASED ON STRUCTURAL TRANSFORMATION MODELS

摘要

An efficient, effective algorithm is presented for matching and recognition of deformed closed contours based on structural features in terms of convex/concave parts and quantized-directional features along contours. In particular, the problems of improving the efficiency of computation, the accuracy of shape classification, and the robustness against noise and global/local shape transformations are addressed. Based on convex/concave structures incorporating quantized-directional features along contours, a compact shape representation with simple, efficient computation is explored so that the contours can be described by a few components with rich features. To cope with structural deformations caused by noise, scales of observation, and global/local deformations, shape transformation models are introduced for editing features, and an efficient algorithm is developed for finding an optimal correspondence of features between the two contours so that total cost for the editing can be minimized. Based on the correspondence of features, a contour is transformed by an affine transformation so that the contour can be as close to the other as possible. The distance between the two contours is calculated by comparing the normalized contour with the other in terms of some distance metric. Furthermore, the effectiveness of the proposed algorithm is validated by systematically designed experiments with a large number of testing sample data.