Hand written zip code recognition by back propagation neural network

A technique of recognizing zip codes by using the back-propagation neural network (BPN) with different structures is presented. The differences between the structures are the number of units in the hidden layers and the number of hidden layers. One of the structures mimics the Neocognitron structure. There are two hidden layers in this structure. The weights between the hidden layers are fixed like the fixed weights between the S and C layers of the Neocognitron. The zip code images were retrieved from the database supplied by the Office of Advanced Technology, US Postal Services. Therefore, they represent a set of totally unconstrained handwritten numerals from the real world. Preprocessing consisting of segmentation and normalization to a standard size was carried out. All the segmented images were normalized to size 19×19 by the averaging method. The aspect ratio of the original images was kept the same. With proper training of the network, it is shown that the BPN is capable of recognizing zip codes