Research On Mongolian Handwritten Recognition Based On Data Augmentation And Correction Network

Handwriting recognition has always been an important research field in pattern recognition and has attracted much attention from the academic community.Handwriting recognition research has developed from simple isolated word recognition to text line recognition,unconstrained handwriting recognition,document recognition and scene text recognition,etc.However,due to the fact that Mongolian handwriting recognition has just started and relevant research materials are limited,coupled with the characteristics of large vocabulary,free writing and serious character deformation of Mongolian,Mongolian offline handwriting recognition is facing severe challenges.Therefore,this paper takes traditional Mongolian as the research object and carries out handwriting recognition research.In view of the problem of insufficient handwriting data of Mongolian,two data enhancement methods are used to expand the original dataset,and the research of Mongolian handwriting recognition is carried out on this basis.Specifically,the research content includes:(1)In order to address the problem of insufficient Mongolian handwriting dataset,double data enhancement is carried out on the original dataset using the improved Generative Adversarial Network and Elastic Transformation algorithm.Firstly,the improved Generative Adversarial Network is trained on the original dataset to generate handwritten images of five thousand Mongolian words under the condition of full learning.Then,the dataset is further enhanced by elastic transformation algorithm.By these two data enhancement methods,the Mongolian handwriting dataset is fully expanded,increasing the quantity and diversity of the dataset.(2)In view of the characteristics of Mongolian handwriting words of different lengths and distorted writing,a recognition model of end-to-end training is proposed by correcting the handwritten image with a correction network and integrating the correction network with the recognition network.The correction network uses a variant of the spatial transformation network,the thin plate spline(TPS)method.TPS uses a smooth spline interpolation algorithm to interpolate between a set of reference points located,and uses pixel values and connection information to normalize the character region into a predefined rectangle.The fusion of the correction network can improve the generalization ability and recognition accuracy of the model.(3)In order to address the issue of poor robustness of the Mongolian handwriting recognition model,a feature fusion-based improved Vi T model(F-Vi T)was proposed.Extracting features from Mongolian handwriting is difficult due to its unique word formation method,and the feature fusion mechanism can extract Mongolian handwriting features in a more comprehensive and detailed manner.The feature fusion mechanism uses Resnet50 as the backbone network and integrates the feature pyramid network(FPN)to form the feature extraction layer of the recognition network.The handwriting recognition model is based on the basic framework of Vision Transformer(Vi T),which performs multi-head self-attention calculation and feed-forward network processing on the feature vector to obtain more representative feature representation.Finally,the TPS correction network and F-Vi T model were combined to propose the Mongolian handwriting recognition model(TF-Vi T)combining the correction network and the improved Vi T network.Compared to the traditional CRNN model,the TF-Vi T model can better adapt to the Mongolian handwriting dataset,reduce the model complexity,and improve the recognition accuracy.Through the above three aspects,this paper achieves a higher level of Mongolian handwriting recognition.The research on Mongolian handwriting recognition can promote the popularization of intelligent information in minority areas,which is of practical value and has important significance for the inheritance and protection of traditional Mongolian.