FUZZY STATE MACHINES TO RECOGNIZE TOTALLY UNCONSTRUCTED HANDWRITTEN STROKES

An automatic off-line character recognition system for totally unconst rained handwritten strokes is presented. A stroke representation is de veloped and described using five types of feature. Fuzzy state machine s are defined to work as recognizers of strokes. An algorithm to obtai n a deterministic fuzzy state machine from a stroke representation, th at is capable of recognizing that stroke and its variants is presented . An algorithm is developed to merge two fuzzy state machines into one machine. The use of fuzzy machines to recognize strokes is clarified through a recognition algorithm. The learning algorithm is a complex o f the previous algorithms. A set of 20 stroke classes was used in the learning and recognition stages. The system was trained on 5890 unnorm alized strokes written by five writers. The learning stage produced a fuzzy state machine of 2705 states and 8640 arcs. A total of 6865 unno rmalized strokes, written freely by five writers other than the writer s of the learning stage, was used in testing. The recognition, rejecti on and error rates were 94.8%, 1.2% and 4.0%, respectively. The system can be more developed to deal with cursive handwriting.