Research On Chinese Character Recognition Technology Based On Convolutional Neural Network

With the development of the times,the research on the recognition of Chinese characters has a wide range of application prospects in the related fields of document digital retrieval,postal mail sorting,bank check processing,table making,and handwriting-based text input,and has attracted much attention.Traditional research methods mainly study the Chinese characters through the mode of "preprocessing+feature extraction+classifier",but due to the problems of character recognition itself,there are many kinds of characters,more confusing characters,and complicated changes in structure,etc.,the research on character recognition is still There are many difficulties and shortcomings,and the convolutional neural network has the advantages of parameter sharing and strong feature extraction capabilities.The use of convolutional neural network technology can effectively achieve accurate recognition of Chinese characters.This paper mainly studies the Chinese character recognition technology based on convolutional neural network.The research contents are as follows:(1).Aiming at the problems of incomplete character categories and little change of character characteristics in the existing Chinese character database,a method for constructing Chinese character character database based on character encoding is proposed,which can be used for training and testing of neural networks.This method uses the official font to get all the corresponding character information,and uses the character encoding to output the corresponding character picture.In order to expand the data set,different character encoding is used to output the character picture in different fonts,and after the image is zoomed out,the image The original character image is processed by the methods of rotation,convex deformation and ripple distortion to obtain the Chinese character data set.This method can achieve the goal of creating a Chinese character database offline by itself.The created image has a large amount of data,many categories,and various image features.It has good generalization capabilities.(2)Aiming at the problems of poor feature extraction ability and large amount of calculation of existing neural networks,a new convolutional neural network architecture is proposed,which can be used for Chinese character recognition.The network consists of 7 convolutional layers,4 pooling layers,2 batch normalization layers and 1 Softmax regression layer connection,and it is optimized by adding data amplification,batch normalization,RMSprop methods,etc.,which improves the accuracy of network identification.rate.At the same time,for the problem of network overfitting,regularization and Dropout methods are used to effectively prevent network overfitting.By comparing it with classic neural network models such as Alexnet and Lenet,it can be proved that the network architecture proposed in this paper has strong ability of extracting features and high recognition accuracy in Chinese character recognition.(3).To solve the problem of high complexity of existing neural network optimization methods,an asymmetric convolution optimization method for convolutional neural networks is adopted.According to the unique nature of Chinese characters in the process of Chinese character recognition,this method does not change the original neural network architecture,and only replaces part of the convolution layer and the batch normalization layer with a hidden layer composed of asymmetric convolution kernels,using asymmetric convolution The kernel strengthens the weight of the four-neighborhood skeleton information to extract the effective features in the feature image.Through experimental comparison with the original network model,it can be seen that this method is superior to the original convolutional network in terms of recognition accuracy and accuracy variance The model embodies the performance optimization of asymmetric convolutional verification and convolutional neural network in Chinese character recognition.