Handwritten Character Recognition Based On Convolutional Neural Network

With the advent of the era of information and intelligence,it is an inevitable trend and requirement of the development of times to realize automatic recognition and processing of a large amount of information through pattern recognition technology in order to continuously improve the convenience of life and the efficiency of work.This paper studies deeplythe method of handwritten character recognition based on convolutional neural network,a handwritten character recognition method based on Google Inception V2-a model and scale specificationis proposed andis successfully applied to the graphical interface of handwritten character recognition.The main work of this paper includes the following aspects:I.This paper analyses the abnormal data in Nist and Chars 74 K datasets,and deal withthembydeleting,modifying,correcting,on the basis alsomakes a comparative study on the influence of character centralization and data preprocessing on model training.The experimental results show thatabnormal data processing and centralization processing can improve the accuracy of model recognition on the test set;min-max normalization,zero-mean,standard normalization and whitening can solve the problem of model non-convergence,but their effects on model accuracy are almost the same.II.Based on the research and analysis of the LeNet,AlexNet,VggNet16,Google Inception V2 and ResNet50 model,this paper improves the five models respectively.The experimental results show that the accuracy of the improved model on the test set have been improved,and the best test results have been achieved on the improved Google Inception V2-a model.III.Taking the improved Google Inception V2-a model and ResNet50 model as examples,this paper visualizes the features extracted from each convolution layer of convolution neural network by deconvolution method.The experimental results show that the first two layers of the model mainly extract the gray information of characters and the structural features of font,while the latter convolution layers mainly extract the abstract features or advanced features related to font.IV.For the similarity problem between different types of characters,this paper proposes an image processing method based on scale specification on the basis of the research of character similarity merging.The experimental results show that the size specification of Arabic numerals and case letters according to the writing standards can greatly improve the accuracy of the model on the test set,and the recognition rate of 62 types of characters can reach 96.64%.V.For groups with different writing habits or different application scenarios,this paper studies character recognition through transfer learning,andtakes the transfer from Nist data set to Chars 74 K data set as an example to simulate.The experimental results show that the local transfer of the model can improve the convergence effect on the new data set,and provide a solution for solving the problem of changing the application group or scene of the model.