The CNN based Machine-printed traditional Mongolian characters recognition

In this paper, we present an approach to recognize the Machine-printed Mongolian characters by CNN (convolutional neural network). Firstly, a training set of traditional Mongolian characters is collected in advance. There are 85 categories in all and each category is trained and recognized by the CNN. And then, we set a CNN with seven layers. There are three convolution layers, two subsampling layers, one full connected layer and an output layer in the CNN. The experimental results show that the CNN is superior to the BP neural networks for the Machine-printed traditional Mongolian characters. In the BP neural network, the average accuracy is 88.34%. The proposed CNN based method can improve the average accuracy from 88.34% to 99.39% and the improvement rate can be attained to 12.51%.