基于改进LeNet-5网络的数字电表识别  

Digital meter recognition based on improved LeNet-5 network

作　　者：张宁宁 赵明冬[1] 周斌 马金辉 Zhang Ningning;Zhao Mingdong;Zhou Bin;Ma Jinhui(Zhengzhou University of Science and Technology,Zhengzhou 460064,China)

机构地区：[1]郑州科技学院,河南郑州460064

出　　处：《无线互联科技》2023年第11期165-168,共4页Wireless Internet Technology

基　　金：河南省教育厅2022年大学生创新创业训练计划项目,项目名称:基于计算机视觉的特殊场景仪表读数系统,项目编号:202212746003。

摘　　要：目前在水下等特殊应用场景的电表识别研究中,虽然LeNet-5网络表现良好,但仍存在泛化能力不足、鲁棒性较差等问题。为此,文章基于改进LeNet-5网络的数字电表识别方法,通过增加激活离群值去除,利用dropout算法和ReLU激活函数增强神经网络泛化能力与鲁棒性。实验结果表明:改进的LeNet-5网络模型在学习速率为0.1%和迭代次数为600次时,网络精度达到99.42%。该方法具有较强的运算能力和较高的网络识别精度,可满足水下数字电表识别需求。The LeNet-5 network has performed well in current research on the recognition of electricity meters in special applications such as underwater,but still suffers from a lack of generalisation and poor robustness.In this paper,we enhance the generalisation ability and robustness of the neural network by increasing the activation outliers and using dropout algorithm and ReLU activation function.Experimental results show that the improved LeNet-5 network model achieves 99.42%network accuracy at a learning rate of 0.1%and a number of 600 iterations.The method has strong computing power and high network recognition accuracy,which can meet the needs of underwater digital electricity meter recognition.

关 键 词：数字识别 改进LeNet-5网络 dropout算法 特征提取 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]