Dynamically connected neural network for character recognition

A two level dynamically connected network architecture was developed for handwritten-digit recognition. In the first level a fast preselection was done which reduced the number of possible character classes to a maximum number of three, and then the input image was propagated through several dynamically chosen precise exclusion networks to confirm or cancel the result of the preselection. The network was trained over a few hundred sample patterns written by the authors, while it was tested over 2000 digits from 200 different persons. The network had no substitution error, and the rejection rate was 5%.<
>