Printed Uyghur Character Recognition Based On CNN

Printed Uyghur character recognition is one of the most important research fields in the field of text recognition in China.The printed Uyghur character recognition technology based on statistical machine learning method needs manually extract features,which include eight direction code feature,gradient feature,histogram feature,width height ratio feature,character template feature and so on.The convolutional neural network with deep learning method can automatically extract features from character images by convolution layers.With training model,the process of feature extraction can be optimized continuously,and the feature extracted by the convolution layers can represent the character images better and can get a high recognition accuracy.Therefore,this paper proposed a method that use the convolutional neural network to recognize printed Uyghur characters.In order to verify the influence of the volume of convolution layer on the character recognition model,one convolution layer,two convolution layers,three convolution layers and four convolution layers are tested respectively,and the features extracted by the four convolution layers can well represent the character images.At the same time,a comparative experiment on models with different learning rates is carried out.It is found that,the model with less number of classifications,an excessive learning rate can easily cause gradient explosion,which leads to the model can not learn the character features,but the small one will get better training results.In the experimental environment,the recognition accuracy of the printed Uyghur character recognition model based on the convolutional neural network can reach 99.63%on the 49 thousands character test data set.