基于卷积神经网络的手写数字识别

Handwritten Numeral Recognition Based on Convolution Neural Network

在LeNet-5模型的基础上,改进了卷积神经网络模型,对改进后的模型及网络训练过程进行了介绍,推导了网络模型训练过程中涉及到的前向和反向传播算法。将改进的模型在MNIST字符库上进行实验,分析了卷积层不同滤波器数量、每批数量、网络学习率等参数对最终识别性能的影响,并与传统识别方法进行对比分析。结果表明:改进后的网络结构简单,预处理工作量少,可扩展性强,识别速度快,具有较高的识别率,能有效防止网络出现过拟合现象,在识别性能上明显优于传统方法。

The convolution neural network model is improved on the basis of LeNet-5 model. The improved model and the network training process are introduced, and forward and back propagation algorithms of network model in the process of training are deduced. The improved model is tested on the MNIST character library, and the effects of different filter number at the convolution layer, quantity of each batch and network learning ratio on the performance of the final recognition are analyzed. Meanwhile, and the traditional identification methods are compared with the recognition method in this paper. The experimental results show that the improved network structure is simple, with small workload of pretreatment, strong extensibility, fast recognition and high recognition rate. It can effectively prevent the network over-fitting phenomenon. The recognition performance is significantly superior to traditional methods.