Fast Handwritten Numeral Recognition Algorithm Based On Residual Network

Handwritten digital recognition technology is widely used in the Internet,banking,education and other industries,and Convolutional Neural Network is one of the important algorithms for handwritten digital recognition.With the continuous evolution of various kinds of Convolutional Neural Networks,the accuracy of handwritten digital recognition is increasingly improved,and the focus of people's improvement on it is gradually changed from "improving the accuracy" to "accelerating the convergence speed of model training".This paper selects the Residual Network as the research direction to improve the Residual Network from the two aspects of network structure and training technology to improve the convergence speed of the network.Firstly,this paper introduces the classical Convolutional Neural Network LeNet,GoogLeNet with multi-channel idea and Residual Network,SqueezeNet,DenseNet,mainly including the introduction of the model structure,principle and advantages of several networks.Subsequently,this paper introduces the classical multi-channel network model Inception-ResNet-v2.In order to improve the convergence speed of the network,an improved multi-channel structure is added to the Residual Network.Different from the original Inception-ResNet-v2 network,which uses parallel connection completely,the improved multi-channel Residual Network is based on the original network,in which a1-by-1 convolution layer is connected in parallel for each convolutional layer to form a multi-channel convolution module,and residual connection is added after each two modules are connected in series.Meanwhile,in order to speed up the convergence of the network,the Dropout technology commonly used is abandoned to improve the network.Then the reasons and advantages of the model improvement are analyzed.Finally,the network model is optimized in training technology.Firstly,the Batch Normalization technology applied to GoogLeNet model is applied to the improved network in this paper to improve the network convergence speed and generalization ability.Then the Adam algorithm used in AlexNet model is used in the improved network in this paper to reduce the amount of computation in the network back propagation.In order tofurther improve the convergence speed of the network,this paper uses the strategy of SVM to process outlier samples to propose a multi-dimensional plane Output Approximation Algorithm based on supervised learning,and analyzes its principle with the method of multi-dimensional plane graphic.The algorithm approximates the test sample to the ideal sample,and there are three key threshold parameters.One is the "threshold of test accuracy".When the test accuracy is higher than the threshold,the network is considered to have a good generalization ability,and the Output Approximation Algorithm is used.The second is the "Euclidean distance threshold".When the distance between a test sample and the ideal sample is greater than the threshold,the sample is considered to be unrepresentative due to excessive noise,and the Output Approximation Algorithm is used to push it to the ideal sample.The third is the "approach distance",which represents the moving distance of the outlier sample.This paper takes MNIST and MNIST after data enhancement processing as data sets,conducts many experiments and draws a conclusion: compared with the traditional Residual Network,the improved multi-channel Residual Network in this paper Can make the network training times from 1000 times to 800 times under the same network complexity.The Output Approximation Algorithm proposed in this paper can make the network converge 100 times in advance without reducing the recognition accuracy.The Euclidean distance threshold is set to 0.3,and the approximation distance is set to 0.01,which can make the network convergence speed reach the fastest.After 70 times of training,the accuracy can converge to more than 97%.