Computer recognition of hand-written characters using the distance transform

A new statistical classifier for hand-written character recognition is presented. The system features a preprocessing phase for image normalisation and a distance transform applied to the normalised image, which converts a B/W picture into a grey scale image. A k-nearest-neighbour classifier follows, based on the distance transform and a suitable metric. The system has an accuracy of 98.96% when applied to the US Post Office zip code database at 0.98% error rate.