Online Chinese Handwriting Character Recognition System Based On Convolutional Neural Network

In handwriting recognition system, the features of online handwriting characters are extracted based on their strokes, orders and track sequences. For Chinese handwriting characters, their structures are complex and there are many similar characters. On the other side, there are many variants even for the same character due to different writing habits of writers, which makes the extracting of stable features for Chinese handwriting characters very difficult. Convolutional neural network(CNN) can automatically extract features from original input data without any extra operations, which greatly reduce errors caused by manually constructed feature set. Further more, convolutional neural networks can execute a convolution operation on local visual domain, thus it may generate more representative characteristics for online handwriting characters based on local features. Therefore, it is possible to apply the CNN into online handwriting Chinese character recognition task.In this paper, we proposed an online Chinese handwriting recognition system based on CNN. Our system extracts features of nature strokes. The proposed system consists of three stages: firstly, we need to regularize the information of points and use necessary interpolation operation in order to complete the normalization of data.In the procedure of the preprocessing, we make use of the linear regularization and linear interpolation methods. Secondly, the CNN is trained with the combination of traditional eight-direction features, in which the features generated by convolutional neural network are adjusted according to eight-direction features. In this way, we can get the convolutional features that are quite different from traditional feature.Finally, we combine the convolutional and eight-direction features with hidden layer and clssify Chinese handwriting characters based on the model.The training data sets used in this paper include SCUT-COUCH2009 collected by the South China University, HIT-OR3 C collected by Harbin Institute of Technology Shen Zhen Graduate School, and the CASIA-OLHWDB1.0&1.1. The testing data is chosen from the ICDAR online Chinese handwriting competition.Evaluation results show that our system reaches a satisfying result on the testing data set. The best result of our system on the first candidate is 97.25%, which isclose to the best record of 97.39%.