摘要
在卷积神经网络的基础上改进了LeNet-5模型,建立了更适合于手写数字识别的神经网络模型,并对改进后的模型及网络训练识别过程进行了详细介绍。将改进后的模型用MNIST字符数据库进行验证,分析了不同卷积层特征图数量、每批次训练数等参数对最终识别性能的影响,并与几种常用识别方法进行比对。通过结果可看出,改进后的新型网络结构简单,识别度高,识别速度快,具有鲁棒性好,泛化能力强等优点。说明改进后的神经网络模型对手手写数字具有很好的识别性能,能满足实际应用需求。
Based on the convolution neural network, the LeNet-5 model is improved, and the neural network model which is more suitable for handwritten numeral recognition is established. The improved model and the network training recognition process are introduced in detail. The improved model is validated by MNIST character database, and the influence of parameters such as the number of different volume maps and the number of training per batch on the final recognition performance is analyzed and compared with several commonly used identification methods. Through the results can be seen, the improved new network structure is simple, high recognition, recognition speed, with good robustness, generalization ability and so on. It shows that the improved neural network model has a good recognition performance for handwritten numerals, which can meet the practical application requirements.
出处
《信息通信》
2018年第1期109-112,共4页
Information & Communications
