Abstract: A method for recognition of an unconstrained written character image employs at least one non-character recognizer in addition to character recognizers for each candidate character. The input image is rejected if any of the non-character recognizers produces a match. The input image is recognized only if a) a character recognizer produces a match, and b) none of the non-character recognizers produces a match. The non-character recognizers are preferably trained on a) images which are not recognized by the set of character recognizers, B) images which are misclassified by at least one of the set of character recognizers result, and c) incorrectly segmented images. In the preferred embodiment of this invention individual segmented digit images are recognized for automatic sorting of mail by ZIP code. In the preferred embodiment plural character recognizers and at least one non-character recognizer are embodied in a neural network using a place encoding output.