Research On Handwritten Mongolian Characters Recognition Based On Generative Adversarial Networks

With the rapid development of computer technology,human beings have entered the era of informatization,big data and cloud computing.As a carrier of information,paper is currently unable to meet the needs of people to save,learn,edit,and study and inherit instant information.With the popular use of electronic devices,especially smartphones with camera functions,pictures are gradually becoming the main carrier of information.Image text recognition can identify batch text images as editable text,thereby further promoting the process of informatization,and achieving better preservation,analysis,and research of related documents.This is true for Chinese characters,and the same is true for Mongolian.With the rise of deep learning,the current mainstream network models for image text recognition include Convolutional Neural Network(CNN),Recurrent Neural Network(RNN)and their variants.Based on the analysis of image text recognition methods and deep learning network models,this thesis conducts in-depth research on how to expand a small-scale handwritten text dataset and how to complete incomplete images and remove handwriting styles for text recognition models.The research work in this thesis is as follows:(1)Data augmentation method introducing fusion edge attentionAt present,there is a problem of poor scale and diversity in the handwritten Mongolian dataset.Aiming at this problem,sorting the existing data enhancement methods and related technologies,and analyzing the advantages and disadvantages of data enhancement methods in the task of handwriting Mongolian characters.Based on the idea that the conditional generation adversarial network can generate the direction of the image,the fusion edge attention data enhancement model is proposed.The edge information of the character is used as a constraint,and the edge attention mechanism and feature diversity regular terms are introduced at the same time,which makes the generated data more sensitive to the change of the edge of the character and the generated sample more diverse.(2)Propose a complementary networkThe handwritten Mongolian incomplete image will cause recognition error or unrecognizable situation during recognition.In response to this situation,inspired by the data enhancement method,this thesis proposes a handwritten text completion network based on generative adversarial network fusion image reconstruction ideas.The completion network uses a residual neural network to minimize the consumption of original information during the completion process,and the completed image is closer to the real image.(3)Propose a normative identification networkIn the image recognition task,the recognition accuracy of handwritten Mongolian characters is significantly reduced due to the different writing characteristics of individuals.Aiming at this problem,inspired by the image translation method,in order to weaken the handwriting style features and use the content features as the main basis for recognition,this thesis proposes a handwritten text specification recognition network based on generative adversarial network fusion of image translation ideas.In the canonical recognition network,the content feature of the real image is used as input through the generator of the generative adversarial network to obtain an image with weakened handwriting style.Improve the final recognition accuracy.