Research On Offline Handwritten Chinese Character Recognition Based On Deep Learning

Offline handwritten Chinese characters recognition is an important research field in pattern recognition.Due to the large Chinese character category,more similar characters,and the randomness of the writer's strokes,offline handwritten Chinese character recognition is extremely challenging in the field of pattern recognition.In recent years,deep learning has made breakthroughs in the fields of character recognition,image classification,target tracking and detection,and even completely replaced the traditional pattern recognition method.Therefore,this paper uses the deep learning method to recognize 3,755 offline handwritten Chinese characters provided by the Chinese Academy of Sciences Institute of Automation(CASIA).The main research contents of this paper are as follows:1.The original data set was parsed and the extracted character images were normalized.2.The Long Short-Term Memory(LSTM),Bidirectional Long Short-Term Memory(Bi-LSTM)and VGG16 were used and the preliminary recognition of offline handwritten Chinese characters was carried out.3.Then,the network structure CRNN combined with convolutional neural network and recurrent neural network is adopted(that is,the feature image is extracted first by convolutional neural network,and then the feature extracted by CNN is used as the input of Bi-LSTM of the recurrent neural network)The data set was further recognized and the recognition rate was greatly improved compared with the experimental results of the first three classical neural networks.4.The convolutional neural network HCCR-Net suitable for the data set is repeatedly constructed according to the recognized character features,and satisfactory results are obtained.5.Based on the previous research,the ensemble learning method(voting method)was used to vote on the test set of the three trained network models Bi-LSTM,CRNN and HCCR-Net.For the determination of the weight of the weighted voting method,ths paper proposes the use of the analytic hierarchy process(AHP)and combines some commonly used evaluation indicators of the network model to determine the reasonable weight of each network model.Finally,this paper comprehensively analyzes the test results of all networks.The recognition rates of LSTM,Bi-LSTM,VGG16,CRNN and HCCR-Net on the CASIA-HWDB1.1 dataset are 78.06%,85.25%,87.62%,92.67%and 94.58%respectively.The recognition results of the relative majority voting method based on class markers and class probabilities are 94.82%and 95.41%,respectively;the weighted voting methods based on class markers and class probabilities are 96.06%and 96.33%,respectively.Compared with the experimental results of several separate network models,the results are greatly improved.