Induction of Fuzzy Rules for Character Recognition
by
José Manuel Fonseca
Universidade Nova de Lisboa
Coauthors: Nuno Miguel Rodrigues Rita A. Ribeiro

The use of fuzzy classifiers for Optical Character Recognition (OCR) is appealing but the main obstacle for its development is the enormous number of rules necessary when we have a considerable number of different objects. This is our case since Portuguese cars licence plates can have combinations of 10 digits divided in two sets of two numbers and 23 characters of the alphabet divided in combinations of 2. In this paper we describe a methodology for automatic generation of fuzzy classification rules using decision trees classifiers. The objective is to produce a fuzzy classifier able to correctly classify isolated numbers and letters extracted from licence plate photographs with varying luminosity and different angles of capturing. Using the proposed methodology only the membership functions must be defined by the user. Based on these membership functions and using a set of examples as training set the system automatically generates the fuzzy classification rules that, according to the tests performed, are able to produce very promising classification results. Besides the explanation of the adopted methodology we also present the results obtained by the resulting fuzzy classifier over 380 examples of characters extracted from licence plate images.

Date received: July 13, 2006