On-line handwritten symbol recognition, using an ART based neural network hierarchy

A neural hierarchy is proposed for the recognition of on-line handwritten alphanumeric and mathematical symbols. The neural hierarchy forms part of a mathematical editor which uses handwriting as the principal means of man-machine interface. The symbols are considered as sequences of strokes which are in turn represented by a vector of the stroke curvature. An adaptive resonance theory (ART)-2 module is used for the unsupervised classification of the normalized strokes, while a recently proposed network is used for the acquisition of a spatial pattern. It efficiently represents the sequence of the eventually repeated strokes. An analog ARTMAP module is employed in order to classify the symbols and assign the appropriate code and name to them. Experimental results are presented which confirm the efficient performance of the neural architecture, especially in comparison to the state-of-the-art classical elastic matching algorithm