Research On Handwritten Numeral Recognition Based On Deep Residual Network

The 21 st century is an era of big data.Digitized information spreads around the world through computers.Traditional computer data entry depends on manual work which is low efficiency and high cost.To organically combine computer office with data entry,handwritten numeral recognition technology comes into being.However,existing handwritten numeral recognition has many problems to be solved due to its own structural characteristics.So,it is still worthy for us to research and explore in both the technical field and the application field.Deep Learning is the research hotspot in the field of machine Learning.It is a computing model with multiple hidden processing layers.The multi-abstraction layer representation of data can be acquired by learning the characteristics.In recent years,the emergence of large-scale data training sets and the development of high-performance computer hardware have played a key role in deep learning.And then deep learning has been applied to many fields such as computer vision and image processing,natural language processor,voice recognition and so on.In this thesis,the neural network theory and related optimization algorithm are introduced first.The main contents include neuron model,multilayer perceptron,Back Propagation,local perception,sampling,weight sharing,convolution operation,activation function,pooling,Softmax regression,batch normalization,regularization,etc.Then the problem of overfitting is analyzed when the deep neural network model is used to study handwritten numeral recognition,and the solution is proposed.The data augmentation method is used to solve the problem that small data sets can easily lead to overfitting.In this solution the correlation theory of generative adversarial networks(GAN)and discrimination network is introduced,including the generator and the discriminator.Then on the basis of this a deep convolution generative adversarial network is designed to complete handwritten data augmentation.Finally,the performance of this method is compared with the traditional method by experiments.The experimental results show that the performance of the generated data is better than the traditional method in improving the model performance,and the method of the data augmentation effectively enhances the model performance.The thesis analyzes the gradient vanishing and the gradient exploding when using deep neural network model for handwritten numeral recognition,and the deep residual neural network is used to solve this problem.Specifically,this thesis introduces relevant theorization about residual network which includes the short connection and the convolution layer bifurcation.Then the variant residual module is added to the neural network to realize the handwritten numeral recognition training.By increasing network width,adding batch normalization and adding Dropout layer,the gradient vanishing and the gradient exploding can be effectively avoided and the learning efficiency of the model can be improved.After that,a recognition model with better performance is determined by analyzing the influence of Batch Size,Learning Rate and Dropout on the network performance.Compared with the traditional convolutional neural network,this model has higher recognition rate and stronger robustness.Finally,a simple handwritten numeral recognition system is designed.The system can carry out three functions as data input,neural network forecast and recognition results output.The experiment shows that this handwritten numeral recognition system has a high recognition rate.