Research And Application Of Arabic Handwritten Character Recognition Method Based On Deep Learning

With the rapid development of information society,pattern recognition has been applied in various fields and achieved remarkable results.Among them,handwritten character recognition has attracted extensive attention of scholars.Handwritten character recognition is an important research field of optical character recognition(OCR).With the continuous application of character recognition in people's life,people put forward higher requirements for the effect of character recognition.In order to achieve more efficient character recognition,improve the accuracy and speed of recognition algorithm,this thesis studies two kinds of deep learning algorithms for Arabic character recognition: Convolution neural network(CNN)algorithm and depth residual network algorithm.Finally,the Arabic handwritten character recognition system based on B/S architecture and deep learning algorithm is implemented.The main research work is as follows:1.In order to improve the recognition accuracy of Arabic handwritten characters,a recognition method based on depth binary feature is proposed.This method uses sign function to generate binary image features in the full connection layer of CNN.In order to solve the problem of gradient disappearance in the back propagation of sign function,a continuous gradient function is co nstructed for sign function,so that CNN can carry out back propagation smoothly.The experimental results show that th e proposed method makes the handwritten character features more concise and easy to distinguish,and the recognition rate can reach 95.15%,which is better than the traditional deep learning algorithm.2.In order to improve the recognition accuracy,Firstly,a 50 layer residual network model(ResNet)is designed to recognize Arabic handwritten characters.The recognition accuracy is 95.51%.Secondly,in order to further improve the recognition accuracy,AM-SoftMax classification loss function is used.The experimental results show that the recognition accuracy of the proposed method is 96.72% compared with the existing methods.The depth ResNet can extract more efficient features and improve the recognition accuracy.At the same time,AM-SoftMax is an improvement of the traditional SoftMax loss function,which makes the classification samples aggregate within the class and disperse among the classes,and further improves the recognition accuracy.3.In order to verify the correctness of the proposed algorithm,a simple Arabic handwritten character recognition system is implemented by using the method and B/S architecture proposed in Chapter 3.Firstly,users need to recognize or train Arabic handwritten characters when drawing through HTML canvas.Secondly,the client converts the Arabic characters collected and drawn into an array,and passes them to the server as predicted characters.At t he service end,the training or prediction request is made through API call to CNN_Binary module.Finally,the trained CNN_Binary weights are saved in the file,and the trained weights are loaded to complete the prediction.The implementation results show that the system has high recognition accuracy.