Research On Improved Algorithm And System Of Handwritten Digit Recognition Based On Neural Network

Handwritten digit recognition is a basic image classification problem with high practical value.It has wide application prospects in the fields of accounting,postal services,and finance.The use of computers to complete handwritten digit recognition will greatly simplify people's lives and save a lot of human recognition costs,which has certain practical significance.There are mainly two methods to deal with image classification problems.The first is the traditional method.In the traditional method,the features of the image are first manually designed and extracted,and then the extracted features are sent to a classifier for classification.The second method is an end-to-end image classification method based on a neural network.With the continuous improvement of computer storage and computing capabilities,neural network technology is in full swing,and image classification methods based on neural networks have gradually become a research hotspot.Owing to the fact that the neural network has the ability to automatically extract complex structural features from a large amount of data,it has exceeded the performance of traditional methods in many problems.Therefore,research on neural network technology is of great significance.This paper uses neural network technology to deal with the problem of handwritten digit recognition.The paper mainly focuses on the following three problems to carry out research and analysis,and gives some improved methods:First,the back propagation algorithm for the classical single hidden layer feed-forward neural network structure has a slow training speed and cannot use unlabeled samples.Combining supervised and unsupervised algorithms,this paper proposes two fast and accurate improved training algorithms for single hidden layer neural network.The first improved algorithm uses restricted Boltzmann machine algorithm,an unsupervised learning algorithm,for weight pre-training.The second improved algorithm introduces a cross-layer connection based on the first improved method,and directly inputs the information of the input layer into the output layer.Experimental results on MNIST,a handwritten digit dataset,show that the two improved algorithms proposed in this paper have better recognition rate and faster training speed than traditional single hidden layer neural network trained with back propagation algorithms.Then,although the traditional convolutional neural network has a certain degree of translation invariance,it cannot solve problems such as rotation and scale invariability.This paper combines the spatial transformer network based on affine transformation and the idea of densely connected networks,and proposes an improved method.The experimental results on CMNIST,a cluttered handwritten digit dataset,show that this improved method can automatically locate the relevant digit part of the cluttered handwritten digit picture and correct the figure to normal pose.In addition,compared with the traditional neural network structure,the dense connection method also has better classification accuracy in this problem,which improves the invariance of the traditional convolutional neural network to a certain extent.Finally,based on Flask framework and Keras neural network framework in Python,this paper designs and implements a handwritten digit recognition system based on convolutional neural network model.The system includes data input,data preprocessing,neural network model prediction and result output modules.Experiments show that our handwritten digit recognition system can quickly recognize the user-entered handwritten numbers and has a good recognition effect.