Research And Application Of Offline Handwritten Chinese Character Recognition Based On Deep Learning

Handwritten Chinese Character Recognition(HCCR)is an important form of humancomputer interaction,which has high practical value in bill processing,document entry and other fields.However,due to the large number of Chinese characters and different writing styles of each person,the accuracy of offline handwritten Chinese characters is not good enough.In recent years,Deep Learning has made a breakthrough in image recognition,object detection and other fields,and has gradually become a common method in Pattern Recognition.Based on the method of Deep Learning,this paper studies the off-line handwritten Chinese character recognition by using the feature extraction technology,which improves the recognition accuracy.The main research work is as follows:(1)Follow the specific situation of CASIA-HWDB1.1 dataset,several Convolutional Neural Network(CNN)models with different structures are constructed to study the recognition effect of different CNN models on offline handwritten Chinese characters based on Alex Net,VGGNet and Goog Le Net.The experimental results show that the end-to-end CNN model based on VGG-11 structure has the highest recognition accuracy of 94.5%.(2)Based on the VGGNet structure,several CNN models with different depths are designed to study the influence of the depth of CNN on the recognition accuracy.The experimental results show that when the model is less than 11 layers,the deeper the model structure is,the higher the recognition accuracy is.(3)The recognition method of CNN model based on the fusion of Gabor feature and HOG feature is proposed.Although the end-to-end CNN model can achieve a nice results,as a black box,CNN ignores some information in specific fields when receiving the original image input.In this paper,Gabor feature extraction and HOG feature extraction technology in the traditional image processing field are used to improve the training dataset,and the extracted feature map and the original training dataset are fused together as CNN's training data.The experiments show that the recognition accuracy of Gabor extraction and HOG extraction is improved by 0.8% and 0.6% respectively.(4)For the problems such as slow training speed and difficult convergence of CNN model,this paper studies the acceleration effect of fine-tuning technology based on Transfer Learning and Batch Normalization algorithm on model training,and analyzes the influence of different dropout probability on model's accuracy.At last,a testing system for off-line handwritten Chinese character recognition is implemented based on Tensorflow framework.