Research And Application Of Handwritten Chinese Character Recognition Algorithm Based On Swin Transformer And CNN

Since Chinese characters are characterized by many categories,complex and diverse strokes,and many similar characters,even the same Chinese character can vary greatly from person to person,which make it difficult to recognize handwritten Chinese characters.Early structure-based and statistical feature extraction methods have some advantages,but they have the disadvantages of requiring large computational resources,low accuracy and long time compared with deep learning.Most of the existing methods focus on using Convolutional Neural Network(CNN)for feature extraction and classification of Chinese characters,which has two advantages of local invariance and translational invariance,but also has the problems of limited perceptual field and "network degradation".The Swin Transformer network,based on the attention mechanism,is able to establish the association between image features through its attention mechanism,which allows the neural network to recognize the results not only based on a single feature but also through the relationship between features,and can increase the recognition accuracy for certain characters with similar shapes and complex strokes.The shortcoming of the Swin Transformer network is that it lacks the inductive bias of CNN,so combining the advantages of CNN and Swin Transformer to study a network structure suitable for handwritten Chinese character recognition is the focus of this paper.In this paper,we study the offline handwritten Chinese character recognition algorithm based on the traditional CNN network model and the Swin Transformer network model,fuse the advantages of the two networks to design a network structure suitable for handwritten Chinese character recognition and apply the studied algorithm to the field of calligraphy teaching,and develop a calligraphy teaching system.The main research contents are as follows:(1)The network models of Swin Transformer and CNN are trained on the offline handwritten Chinese character data set,and the ST-CNN structure combining CNN and Swin Transformer is designed by comparing and analyzing the experimental results and the respective advantages of Swin Transformer and CNN to complement each other.In order to solve the "degradation" problem of neural network,we adopt the idea of Res Net(residual network)in designing CNN,and further compare the experimental analysis among CNN,Transformer and ST-CNN.The experimental results show that the ST-CNN network based on the combination of CNN and Transformer improves the convergence speed by about 10% compared with CNN,30% compared with Swin Transformer,and the accuracy rate improves by about 0.5% compared with Swin Transformer,and the algorithm is compared with other common network structures.(2)The trained ST-CNN network was used as the main algorithm for the development of the calligraphy teaching system,which mainly includes the Android application,the server and the study of network communication using Socket,and finally the interaction of information between the server side and the Android application side was realized.The final test showed that the system can identify the specification or not of the font writing and give suggestions for writing improvement to achieve the purpose of calligraphy teaching.