Chinese Handwriting Recognition Method Using Deep Learning

As the premise and foundation of the Chinese documents digitization,Handwritten Chinese recognition is of important values in historical document recognition,transcription of handwritten notes,etc.Chinese handwriting recognition is a challenging problem due to its free writing forms,complex structures,strong similarity,the large number of dataset,and the lacking of standardization.Firstly,this thesis analyzes the research background and significance of Chinese handwriting recognition,and summarizes the research status at home and abroad.Then the basic theory of deep learning and the main components of convolutional neural network(CNN)are introduced,as well as the commonly-used models.Secondly,an improved Le Net-II network model is proposed based on the traditional Le Net-5 one.Using the improved inception module and dilated convolution,a collateral dual CNN structure is designed.The two branches are used to extract the Chinese handwriting images’ features in different scales,and then to obtain the feature images in multiple scales.After feature fusion,the diversity of feature images is enriched,thus the recognition accuracy is improved.By training the classic Chinese handwriting dataset with the proposed network model,3755 handwritten Chinese characters and relative Chinese documents are recognized.Compared with the classical Le Net-5 model and traditional algorithms,experimental results demonstrate that the improved model performs well in the aspect of both convergence speed and recognition accuracy.Besides,the average recognition accuracy on 6 Chinese handwriting documents reaches 97.13%,which is higher than that of human performance.Since the pooling calculation tends to ignore some feature information in dimensionality reduction,a Chinese handwriting recognition method based on attention mechanism is proposed to solve this problem.Based on the traditional CNN model,an AT-CNN network model is proposed.The information interaction between each layer in network is realized using attention mechanism,thus the information loss caused by pooling operations is reduced.Experimental results on the classic HWDB dataset show that the recognition accuracy of 95.05% can be reached using the proposed method.The recognition accuracy is significantly improved compared with that of CNN and other traditional methods.In addition,a Chinese handwriting recognition GUI interface is designed,which is convenient for users to operate.Finally,this thesis summarizes our works and expects the study direction in the future.