Multi-channel handwritten digit recognition using neural networks
Zheru Chi, Zhongkang Lu, Fai-Hung Chan · 2002
Human recognition is much more robust than machine recognition in dealing with rotated and noisy patterns. In this paper, we present a multi-channel neural network (MCNN) approach for handwritten digit recognition based on the human recognition experience in the hope of achieving human-like performance. In this approach, three neural net work modules are trained individually by using three different set of features, intensity-based, rotation invariant, and noise deducted features. The outputs of these three modules are then combined by a combination neural network which is trained separately. Experimental results on a database of 1900 digit patterns written by 190 people show that a recognition rate of 89.5% is obtained on an independent test set that includes both the rotated and noisy data.