Research And Design Of Handwritten Digit Recognition Based On Convolutional Neural Network

As an important branch of artificial intelligence,handwritten digital recognition technology has an extremely wide application prospect.The technology can be applied to hand-written digital automatic recognition and input systems such as financial statements,postal automatic sorting,testing score statistics,bank documents,and financial statistics.Under the background of the global informatization development and the improvement of automation,the demand for handwritten digit recognition technology has become very urgent,and it has important significance and value to develop accurate and efficient identification methods.With the rise of research on image classification by convolutional neural networks,it has brought infinite possibilities for handwritten digit recognition.The unique two-dimensional data processing method of the convolutional neural network and the feature of automatic extraction of image features in the classification recognition can improve the generalization ability and accuracy of the handwritten digit recognition.This thesis relies on the corporate project of the Tsinghua University Soda Research Center Big Data Processing Center.For the problems of handwriting digit recognition,such as inefficient performance and insufficient accuracy,based on the advantages of convolutional neural network in image classification processing,a series of research and design work on handwritten digit recognition has been carried out.Focusing on the design and optimization of the LeNet-5 model of the convolutional neural network,it achieves higher recognition rate and better performance while achieving handwritten digit recognition.For the problems of traditional digital handwriting recognition methods such as feature extraction data redundancy,complex image preprocessing,slow convergence,poor generalization ability,etc.,this thesis proposes the application of convolutional neural network to handwritten digital recognition.The seven-layer structure and working principle of convolutional neural network LeNet-5 model are studied.The forward propagation algorithm,backward propagation algorithm and gradient descent algorithms used in convolution neural network training are determined and deduced.The training parameters and initialization parameters of each layer of the network model are designed.Using the Caffe deep learning framework to build a LeNet-5 model of handwritten digit recognition,the MNIST data sample library was used to train and test the network model.On this basis,handwritten digit recognition based on the convolutional neural network LeNet-5 model is implemented,and the experimental results are analyzed to verify the feasibility of convolutional neural network in handwritten digit recognition.For the problems that the convolutional neural network LeNet-5 model uses in handwritten digit recognition,three optimization optimization methods are proposed in the thesis: 1.To effectively solve the local minimum problem,add a weight attenuation term in the network loss cost function;2.In order to improve the generalization ability and recognition rate of the network and avoid overfitting,the Dropout technology is used to construct the network high-level structure;3.In order to solve the problem of non-convergence and improve the convergence speed when LeNet-5 model is trained and identified,ReLU function is selected to replace the commonly used Sigmoid and other saturation functions as the activation function of each layer output in LeNet-5.This dissertation combines the improvement items together to build a network model with the best performance: LeNet-5+,compare and analyze the recognition results and performance of the LeNet-5 model.Through the experiment,we explored the network misrecognition rate and recognition rate of different models.Iterative error and the effect of network convergence speed.The improved network model increases the recognition rate from 93.61% to 97.30%,the misrecognition rate from 0.96% to 0.75%,the rejection rate from 5.43% to 1.95%,the iteration error from 0.40 to 0.15,and the recognition convergence speed is also improved.Greatly improve.This validates the effectiveness and advantages of the improved network model LeNet-5+ in handwritten digit recognition.At present,the technical achievements studied in this dissertation have been successfully applied to a series of artificial intelligence education products of Chengdu Zhunxing Yunxue Science and Technology Co.,Ltd.,providing intelligent and accurate teaching services such as note recognition and automatic judgment.The research results of this paper have great practical value and significance for the application of handwriting digital recognition technology to artificial intelligence education.