Handwritten numeral recognition via fuzzy logic and local discriminating features

Abstract

This paper describes a system to recognize disconnected handwritten numerals based on the concept of fuzzy logic and discriminating local features extracted from numeral images. Initially, the skeleton of an unknown numeral is obtained and decomposed into several segments called branches. The branches, due to their nature, present fuzzy characteristics in terms of their straightness and orientation. Precisely the three fuzzy sets were defined and used to classify branch segments into straight line segments, parts of circles and circles. The membership grade functions are built for character branches and their values are computed for the sequences of pattern branch features which represent numerals. A numeral image is classified to sequence of branch pattern features with the largest overall membership value. In the case of tie, some local topological features such as the number and the position of end points, intersection points and bend points, are used for the classification.

Work supported by São Paulo State Foudantion - FAPESP