Research Of Offline Handwritten Chinese Recognition

With the development of computer technology,people are more dependent on computers receive and process information.There is still a lot of information stored in paper documents,which converts the information in paper documents into computers,making it easier to store,edit and manage.Offline handwritten Chinese character recognition technology can convert Chinese characters written on paper into Chinese electronic documents that can be edited in computers.It is widely used in the fields of ticket recognition,automatic marking system,manuscript document entry and so on.Therefore,it is of great significance to research offline handwritten Chinese character recognition.At present,there are two main types of offline handwritten Chinese character recognition:one is the recognition of handwritten Chinese single-character;the other is the recognition of handwritten Chinese text line.The difference between them is that text line recognition save the step of single-character segmentation.This thesis uses the deep learning method to recognize offline handwritten Chinese single-character and text line.For the recognition of offline handwritten Chinese single-character,this thesis improves the single-model recognition and multi-model voting recognition methods respectively.For the recognition of single model,in order to extract more comprehensive features and improve recognition accuracy,a method of combining convolutional layer,Inception V1 and residual unit block to form a new structure is proposed,in the open CASIA-HWDB1.0-1.I data set,the character accuracy rate reached 97.07%.Compared with the traditional convolutional neural network recognition method,the character accuracy rate increased by 0.49%.For the recognition of multi-model voting,in order to improve the recognition accuracy,this thesis proposes a multi-type network voting method,which uses VGGNet,GoogLeNet,ResNet three commonly used network voting,in the open CASIA-HWDB1.0-1.1 data set,the character accuracy of these three models is 96.60%,96.52%and 96.89%respectively.The character accuracy rate after voting is increased by 0.40%,which is 0.23%higher than that of the same type of network voting.For the recognition of offline handwritten Chinese text line,this thesis uses Encoder-Decoder model based on Attention mechanism.Compared with MDLSTM-RNN network,it can converge without expanding dataset and pre-training,in the public CASIA-HWDB2.0-2.2 data set,the character accuracy rate reached 94.29%.Since the network structure of this thesis uses two layers of LSTM,the sequence learning ability is strong.Therefore,this thesis proposes a method of adding blank labels between adjacent Chinese characters in text lines to suppress the ability to learn semantic information.And the character accuracy rate increased by 1.47%,compared with the MDLSTM-RNN identification method,the accuracy rate is increased by 12.26%.