This paper presents an on-line hand-printed character recognitionsystem, tested on datasets produced by the UNIPEN project, thusensuring sufficient dataset size, author-independence and a capacityfor objective benchmarking. New preprocessing and segmentationmethods are proposed in order to derive a sequence of strokes foreach character, following suggestions of biological models forhandwriting. Variants of a novel neuro-fuzzy system, FasArt (FuzzyAdaptive System ART-based), are used for both clustering andclassification. The first task assesses the quality of segmentationand feature extraction techniques, together with an analysis ofShannon entropy. Experimental results for classification of thetrain--r01--u02 UNIPEN dataset show real-time performance and arecognition rate of over 85/100, exceeding slightly Fuzzy ARTMAPperformance, and 5 inferior to the rate achieved by humans.
展开▼