Experimental study of a novel neuro--fuzzy system for on-line handwritten UNIPEN digit recognition

摘要

This paper presents an on-line hand-printed character recognition system, tested on datasets produced by the UNIPEN project, thus ensuring sufficient dataset size, author-independence and a capacity for objective benchmarking. New preprocessing and segmentation methods are proposed in order to derive a sequence of strokes for each character, following suggestions of biological models for handwriting. Variants of a novel neuro-fuzzy system, FasArt (Fuzzy Adaptive System ART-based), are used for both clustering and classification. The first task assesses the quality of segmentation and feature extraction techniques, together with an analysis of Shannon entropy. Experimental results for classification of the train--r01--u02 UNIPEN dataset show real-time performance and a recognition rate of over 85/100, exceeding slightly Fuzzy ARTMAP performance, and 5% inferior to the rate achieved by humans.