Boundary-based corner detection using neural networks

In this paper we present a novel boundary-based corner detection approach using artificial neural networks (ANNs). Two neural networks are proposed: one for detecting corner points with high curvature, and the other for detecting tangent points and inflection points that generally have low curvature. For a given boundary point p(i), the first ANN uses the normalized coordinates of points on the forward arm (neighboring points succeeding p(i)) or on the backward arm (neighboring points preceding p(i)) of the point p(i) as the input vector. The output feature of the network is the angle of the forward arm (or backward arm) with respect to the x-axis. The boundary point with sufficiently small angle between the forward and backward arms is identified as a corner. Since the feature points of tangency and inflection have relatively low curvature, the signs of curvature, rather than the magnitude of curvature, for points in the neighborhood of p(i) are used as the input vector to the second ANN. The curvature sign at each boundary point is derived from the outputs of the first ANN. The outputs of the second ANN only respond to the sign patterns of tangent points and inflection points. By using both ANNs, all features of corners, tangent points and inflection points can be extracted from the boundary of any arbitrary shape. Experimental results have shown that the proposed ANNs have good detection and localization for objects in random orientations and with moderate scale changes. Copyright (C) 1996.