A Study On The Methods Of Handwritten Numeral Recongnition Based On Ensemble Convolutional Neural Network

Handwritten digit recognition is an important branch of pattern recognition.It mainly studies how to intelligently recognize Arabic numerals in different scenes through computers.At present,handwritten digit recognition technology is widely used in the Internet,finance,education and other industries.Since the 21st century,with the rapid development of the Internet and big data,and the continuous improvement of global informatization and automation,more and more work is required to recognize handwritten digits,such as bank verification of financial bills,company verification of financial statements,post Automatic sorting and school test score statistics.Traditional manual processing has strong limitations,especially in terms of speed.It is difficult to meet the related requirements.The automatic identification method can not only improve the efficiency of digital identification processing and reduce the consumption of manpower,but also effectively reduce the error rate caused by human operations,and make related work more automated,accurate and efficient.Based on a single convolutional neural network,this paper first studied the handwritten number recognition task,proposed a fusion optimization algorithm,and proved the feasibility of the algorithm in MNIST data set and USPS data set,and then carried out further research in combination with convolutional neural network and ensemble learning to improve the recognition performance.This article first introduces basic theories related to convolutional neural networks,including artificial neural network,neuron model,convolutional layer,pooling layer,fully connected layer,backpropagation algorithm,local receptive field,weight sharing,and loss Function,activation function,overfitting,regularization method,Dropout,etc.Then,the content of ensemble learning is introduced,including ensemble learning methods:Boosting method,Bagging method,and ensemble learning strategies:voting method,average method,learning method.Then a single convolutional neural network for handwriting classification tasks is designed.The ReLU function is used as the activation function to solve the problem of vanishing gradients that is easy to occur during training.Padding operations are used in the convolutional layer and the pooling layer to retain more original sample boundary information.Dropout strategy is used at the fully connected layer to suppress overfitting.This paper also proposes an Adam-SGD fusion optimization algorithm that combines the Adam optimization algorithm and the SGD optimization algorithm.Experiments on the MNIST data set and the USPS data set show that this method can help the network model converge to a good result faster during training.Finally,Combined with convolutional neural network and ensemble learning,a handwritten digit recognition model based on ensemble convolutional neural network was designed.Six convolutional neural networks with different structures were integrated using voting and learning methods.Finally,99.57%and 99.59%classification accuracy were obtained on the MNIST data set,and 98.87%and 99.02%classification accuracy were obtained on the USPS data set.The experimental results show that the integrated model is superior to a single convolutional neural network in recognition performance.