Structure of neural networks for industrial character reader

Abstract

Neural networks to recognize industrial characters are required to achieve high reading reliability. To achieve this high reliability, a method to control the structure of a neural network's hidden layer has been introduced. The method defines the feature extraction functions of neurons in the hidden layer, and preliminary teaching is so constructed that it gives the hidden layer neurons defined properties. After the desired property attached to the hidden layer neurons, the ordinary backpropagation procedure refines the structure of the neural network.<>