Offine Handwritten Chinese Character Recognition Based On Deep Learning

Offline handwritten Chinese character recognition has great application prospects in modern office,finance,intelligence and other fields.However,the offline handwritten Chinese character recognition objects are generally collected by scanners or cameras.These Chinese character images are affected by distortion,fuzzy orientation,limitations of imaging equipment and other reasons,so there are few effective features that can be used.In addition,the characteristics of Chinese characters,such as large number of categories,complex strokes and many similar characters,increase the difficulty of offline handwritten Chinese character recognition.Therefore,offline handwritten Chinese character recognition is not only of great research value in the field of pattern recognition,but also an arduous challenge.With the rapid development of deep learning in image classification,natural language processing and target detection,it has become an important way to solve offline handwritten Chinese character recognition.Based on the study of the existing offline handwritten Chinese character recognition algorithm,aiming at the shortcomings of the current recognition algorithm,this paper improves the offline handwritten Chinese character recognition method based on deep learning from the perspectives of convolution neural network model and interpolation reconstruction of the recognized image,so as to improve the recognition accuracy.The details are as follows:(1)The offline handwritten Chinese character recognition network model is constructed based on the Efficient Net network structure.Aiming at the problem of reducing the recognition accuracy in the Efficient Net network structure,the Efficient Net network model of randomly disordered grouping is proposed to improve the recognition accuracy of offline handwritten Chinese character recognition.By studying the structure of group convolution,the shuffle algorithm is introduced to randomly scramble the channel information in the group,so as to complete the information fusion between different groups.Based on this,a randomly disrupted packet convolution network structure is proposed to replace the mbconv module in the original Efficient Net network structure.The problem of using a large number of point by point convolutions in the structure is improved,which reduces the recognition accuracy and improves the recognition accuracy of Efficient Net network model.(2)A double attention model structure is proposed to improve the feature representation ability of Efficient Net model.In order to solve the problem that theEfficient Net is not able to represent the features of the classification objects in the image classification task and the recognition accuracy is not high,this paper proposes a double attention model structure which is composed of channel attention mechanism and spatial attention mechanism.The double attention model reconstructs the feature map from the channel and space of the input feature map to improve the feature representation ability of the Efficient Net.(3)In order to improve the recognition accuracy of low resolution handwritten Chinese character image,an image interpolation algorithm for texture detail and edge structure preservation is proposed.In this paper,the texture detail and edge structure preserving image interpolation algorithm is proposed.Based on the two variable rational interpolation model,the original interpolation algorithm with fixed spatial distance is improved to the interpolation algorithm with variable spatial distance,which improves the smoothness of the edge region.Through image interpolation and reconstruction,the texture details and edge smoothness of the reconstructed image are improved,and the features of the original image are retained to the maximum extent to improve the recognition accuracy of the image.Experiments show that the network structure designed in this paper has a high recognition accuracy in the process of offline handwritten Chinese character recognition.When the image resolution of the recognized image is low,the texture detail and edge structure preserving image interpolation algorithm proposed in this paper can reconstruct the image into a high resolution image,retain the features of the original image to the maximum extent,and improve the recognition accuracy of the low resolution handwritten Chinese character image.