Experiments with robust similarity measures for OCR

This report describes a system for handprinted character recognition. The goals of this work are to understand better the role of shape, shape representations, and shape similarity in character recognition, and to develop character recognition methods that are more robust to degradations and distortions of character patterns than existing methods. In the system, methods previously used in visual object recognition are used to determine the similarity between input character patterns and stored character prototypes. The shape of character patterns is represented by their boundary, sometimes annotated with information about local orientation and curvature. The effects of different representations of shapes and different similarity measures are evaluated and compared in tests with character patterns from the NIST Special Database 3.