Using normal patterns in handwritten character recognition

This paper presents a system that overcomes the dependence on pattern transformation, like translation, rotation, scaling and further deformations of the input to a recognition system, by reducing the pattern to a normal form. The reduction may be viewed as pre-processing that uses different algorithms to reduce the pattern to normal form at: 0, 1, 2, .., n-level. Our system performs, on patterns representing binary images of characters, the reduction to a normal pattern of level 0, 1 and 2, that in practice correspond, respectively, to character extraction, scaling and rotation until the recovery of a standard condition for these. The patterns so normalised are supplied as input to a recognition system, constituted by a Hintzman neural network, that is a content-addressable-memory, which has well known problems of sensitivity to the input variations