Genetic selection of multilayer neural networks for handwritten digit recognition to aid the blind

Abstract

This research aims to develop character recognition capacity in a system to aid the blind to read. This paper presents a method for selecting the neural network configuration and the training procedures using an augmented set of patterns, to improve the handwritten digit recognition rate. A genetic algorithm is used to search among configurations of two unequal hidden layer networks for feed-forward, fully connected neural networks. Training procedures involving augmented sets of training patterns is produced by two methods: by adding to the original set the four shifted positions about the center, and second, by magnifying +10% and -10% every handwritten digit of the original training set. It is found that the recognition performance not only depends on the architecture but also on the training method. The best recognition rate of 94.2% is obtained in a genetically selected neural network of two unequal hidden layers, and trained with augmented patterns by shifting and magnification.