Research On Ofline Handwritten Chinese Character Recognition Based On Deep Learning

Offline handwritten Chinese character recognition(Offline-HCCR)has a wide range of application prospects and values in such tasks as document digitization,handwritten signature recognition,and natural scene handwritten Chinese character recognition.In the field of handwritten Chinese character recognition,Offline-HCCR has always been a difficult problem.Offline handwritten Chinese characters generally have only twodimensional spatial position information,and the handwriting Chinese character writing style is relatively random,and some Chinese character fonts are more complex and similar.These factors increase the difficulty of handwritten Chinese character recognition.In recent years,deep learning technology has achieved great success in computer vision,natural language processing,speech recognition,and text classification.In the field of OfflineHCCR,there has also been a major breakthrough in the method of deep learning.After studying the shortcomings and deficiencies of the existing offline learning handwritten Chinese character recognition algorithms based on deep learning,the following work and improvements were mainly done:(1)Deep learning based Offline-HCCR usually has a deeper network model,and its model parameter amount and computational complexity are relatively high.When the network model is deployed to devices with limited computing resources and less storage capacity,higher computational complexity makes offline real-time recognition difficult.Larger network models occupy more storage resources and make offline deployment more expensive.Based on the deficiencies of these algorithms,this paper improves the deep convolutional neural network,which makes the network model maintains a high recognition accuracy while greatly reducing the model parameters and computational complexity.Using a convolutional neural network of residual structure can make deeper network models easy to fit,and the computation of convolutions is greatly reduced by the improved depthwiseseparable convolution.Constructing a residual network based on depthwise-separable convolution can effectively reduce the computational complexity and model parameters of the model.The recognition accuracy of the model reaches 96.50%,and the model parameter amount is 20 MB.(1)In some actual scenes,since Chinese characters are written arbitrarily,the characters are easily offset,which increases the difficulty of recognition of handwritten Chinese characters.Before handwriting Chinese characters into the deep learning model for training,by adding a spatial transformation network module to correct the bent handwritten Chinese characters,the input of the deep learning model is a shape-regular handwritten Chinese character,thereby increasing the recognition accuracy and robustness.The spatial transformation network can learn the spatial transformation relationship between the original character image and the shape-regular character image through iterative training.