Robust partial shape classification using invariant breakpoints and dynamic alignment

Abstract

A system for classifying partially occluded noisy shapes in varying positions, orientations, and dimensions is described. An iterative algorithm derived from the colinearity principle is developed to locate invariant breakpoints on a shape contour. The set of invariant breakpoints partitions the contour into a sequence of contour segments. Each contour segment is described by an ordered sequence that represents the Euclidean distance between the pixels in the contour segment and the centroid of the region formed by connecting the end-points of the contour segment by a straight line. Two stages of dynamic alignment with the appropriate constraints are formulated to determine shape similarity. The first stage gives a distance measure of contour segment similarity, and in the second stage, the intersegment distances are used to optimally align contour segments so that a final measure of shape similarity is obtained. The performance of the system is demonstrated by considering the classification of aircraft shapes belonging to four classes and classification results are presented as functions of noise and occlusion levels. The results indicate that reasonable classification is obtained for noisy shapes with 0 to 30% occlusion.