Handwritten digit recognition using combined neural networks

In this paper we propose a solution for integrating two neural networks which are used for handwritten digit recognition. As with the human brain structure, we have chosen to send global information to the first network and to send a sequence of local information to the second network. The first neural network uses digits as images. The second neural network uses digits as contours. One strategy is described for combining these neural networks. We show that we can obtain much better results from a combination of neural network than from individual neural networks.