Deep Learning Based Handwritten Chinese Character Analysis

Handwritten Chinese character recognition has a complex research background and widely used in mail information sorting,handwritten document recognition and electronic device input.In recent years,deep learning based methods have been widely applied in image processing,binging revolutionary breakthroughs and developments for handwritten Chinese character recognition.Compared with traditional methods,high recognition accuracy and end-to-end solution make it become the mainstream research method in this field.However,the proposed methods are mainly based on the existing network structure in related image processing field,without fully utilizing the unique structural features and information of Chinese characters.Therefore,further research is needed for this filed.In this thesis,we proceed from the characteristics of Chinese characters and further explore the feature extraction methods from the radical components.Base on the radical regions,a new convolutional neural network structure framework is proposed to improve the recognition accuracy for handwritten Chinese character recognition.Compared to the existing feature extraction methods,the proposed method can include features of the radical level by modifying the fully connected layer to radical region module and fully connected module.The output of the network could contain both radical level and character level information.In contrast to previous methods,the algorithm takes the features of the character level information as the global information and the features of the radical level as the local information,so that the output features can retain more information.Its characteristics enable the algorithm to further enhance the recognition ability of similar words.Based on this,we propose a new training data enhancement method.Considering that the radical components in most characters have different writing styles,we divide the samples in the database into subgraphs of different radicals according to each type of Chinese character.Then new training samples are generated by randomly recombining them from the candidate subgraphs.In addition,in order to effectively reduce the intra-class distance of the network output in the feature space during training,a center loss function is employed to jointly train the network.Since the proposed CNN structure could retain the feature information from the radical components and the radical components have a close relationship in the phrase,we further use is as a new feature extractor to the input images.After the context of the text line is converted into an image sequence by using the sliding window,the input sequence content is identified in the bidirectional long short-term memory model.In order to verify the effectiveness of the proposed methods in this paper,we tested the handwritten character and text recognition on CASIA-HWDB.The experimental results show that the proposed convolutional neural network and training methods could effectively improve the recognition rate of offline handwritten Chinese character recognition to 97.91%with single model,which could further improve the recognition accuracy among the published works.Based on the proposed CNN structure,the BLSTM model could further improve the recognition accuracy of offline handwritten Chinese text recognition to 90.78%.Compared with other no language model methods,the proposed methods could achieve great performance.