A new shape representation scheme and its application to shape discrimination using a neural network

Abstract

A new method of shape representation and feature extraction is suggested. A shape is approximated by a constant-point polygon in which between any two adjacent vertices of the polygon, the number of points on the contour of the shape is constant. This representation is applicable for both concave and convex shapes and there is no chance of missing any spikes on the boundary. The sequence of the angle of variation between two consecutive line segments is taken as the primary feature (representation of the shape). This sequence is then modelled by an autoregressive (AR) process and the least square error estimate of the AR coefficient vector is used as input to a multilayer perceptron (MLP) network for learning and classification. Robustness of the shape representation scheme and the MLP classifier is also investigated empirically. Adaptive AR modelling is used for estimating the numerically stable and robust coefficient vector.