Learning Shape Features Using a Binary Tree Classifier

A vision system that can be trained to learn objects automatically is under development at Oakland University. It can be thought of as an expert system that learns its rules by examples. This paper describes an algorithm that allows the vision system to learn the shape features of an object for use in a binary tree classifier. At each node of the tree a feature and threshold value are selected by maximizing a modified version of the Kolmogorov-Smirnoff distance based on the training samples. The resulting binary tree classifier is suitable for real-time applications, inasmuch as traversing the tree involves only a simple comparison at each node. An example is given in which shape features of a binary image are calculated using a single-pass shape extractor algorithm.