Optical character recognition using automatically generated Fuzzy classifiers

Character recognition using Fuzzy classifiers has been showing very promising results. However, the definition of the membership functions together with the design of the classification rules is a challenging task even considering just the 10 digits and 23 characters of the Roman alphabet. In this paper we present a solution for the semi-automatic design of a Fuzzy classifier for letters and digits to be applied on the automatic recognition of cars license plates on unstructured conditions. Based on a training set of fuzzified examples of measures, taken from digital images of single characters, the CART algorithm learns the rules that regulate the design of the different characters and generates fuzzy rules that implement the fuzzy classifiers in a completely automatic way. After, a fuzzy inference engine executes the rules to obtain the characters classification. To take advantage of syntactical correction, a hierarchical classifier with two layers of classifiers is proposed: one classifier distinguishes between letters or digits; the second layer classifies either the letters or the digits. The performance achieved by the two-layer classifier is shown and discussed.