A NEURAL-NETWORK CLASSIFIER FOR OCR USING STRUCTURAL DESCRIPTIONS

We present a method for character recognition especially designed for the case in which the shapes of characters belonging to the same class vary greatly, as it happens with unconstrained hand-printed character s and omnifont printed characters. The most distinctive feature of the method is the use of a special kind of structural description of char acter shape in connection with a neural network classifier. An origina l technique is used to achieve the best trade-off between reject and m isclassification rates. Experimental results on databases of both hand -printed and printed characters are illustrated.