Research On Chinese Character Recognition Method Based On Deep Learning

Characters exist in various forms in the process of human life,and play a very important role in the economic,scientific and technological exchanges and development of the whole society,which is also an indispensable and important part of human life at present.In the face of these complex and huge amount of data,how to extract the valuable information from constructing different models or algorithms removing a large amount of spamming,has always been an important problem in research.The tradition Chinese character recognition is mainly limited to feature extraction,while convolution neural network model in deep learning has the advantage of being able to extract features automatically from a large number of samples.This topic has carried out a certain research through a combination of convolution neural network model and the traditional application of pattern recognition of Chinese character recognition and the main research work of the paper is in the following aspects:(1)Studied the traditional Chinese image recognition methods.Support vector machine(SVM)and decision tree set were used as classifiers,and three different pretreatments,namely square feature,gabor-based feature and elastic mesh feature,were used for the comparison experiment.Experimental results show that the morphological transformation of the data set preprocessing method combined with elastic mesh features,can obtain better recognition effect.(2)The model of convolutional neural network based on attention mechanism was studied.In order to effectively solve the problem of loss in the training process caused by small differences in similar Chinese characters,the attention level and AlexNet convolution level were effectively combined to further improve the weight of small differences in Chinese images and reduce the influence of convolution level on lost information.(3)In the process of experimenting with Chinese character recognition,data enhancement was studied,and the results showed that after reasonable data enhancement and expansion,the problem of over-fitting of deep convolutional neural network could be effectively avoided.