Classification of printed characters using multi-layer feedforward neural networks

Multilayer feedforward neural networks for bit-map classification of 95 printed characters are evaluated. The error backpropagation algorithm performance has been investigated for different learning parameters and architectures. Learning profiles are compared, and the most suitable learning conditions are outlined. The best classification results have been obtained with a two hidden layer network using about 70 hidden units per layer. Learning of this architecture has been quick and reliable; this size of layers presumably provides excellent redundancy. It has also been found that this size could be reduced to approximately 25 without major deterioration of classification. Training of networks with too many or too few hidden units has resulted in slow learning with a somewhat large number of decision errors