摘要
针对输电线路机巡影像缺陷识别中低漏报率的需求,提出了一种基于组合式深度目标检测框架的输电线路低漏报率缺陷识别方法。该方法首先利用典型目标检测算法在输电线路巡检图像数据集上进行训练,得到输电线路设备缺陷的特征提取网络;随后引入位置随机分布函数改进目标预测的方式,并利用自适应非极大值抑制判别器,对2个网络的特征提取结果进行自适应融合,最终得到巡检图片中缺陷的类型和位置。测试结果表明,该方法能够有效降低巡检图像缺陷识别的漏报率,采用该方法得到的主要缺陷的平均漏报率远低于其他深度学习模型,可同时实现多类缺陷的检测,能有效促进输电线路常规巡检中缺陷自动识别的应用和推广。
In order to solve the problem of underreporting in the inspection of defects by transmission line machine,we proposed a defect recognition method of power transmission with a low underreporting rate based on combined deep target detection framework.In this method,first,two target detection algorithms were used to train on the transmission line inspection image data set to obtain the feature extraction network for transmission line equipment defects.Then,the location random distribution function was introduced to improve the method of target prediction,and the adaptive non-maximum suppression discriminator was used to adaptively fuse the feature extraction results of the two networks to finally obtain the type and location of the defects in the inspection picture.The test results show that this method can be adopted to effectively reduce the underreporting rate of inspection image defect recognition.The average underreporting rate of major defects is far lower than that obtained by other deep learning models;meanwhile,the detection of many kinds of defects can be simultaneously realized,which can effectively promote the application and promotion of automatic defect identification in the routine inspection of transmission lines.
作者
罗鹏
王波
马恒瑞
马富齐
王红霞
朱丹蕾
LUO Peng;WANG Bo;MA Hengrui;MA Fuqi;WANG Hongxia;ZHU Danlei(School of Electrical Engineering and Automation,Wuhan University,Wuhan 430072,China;Tus-Institute for Renewable Energy,Qinghai University,Xining 810016,China;Central Southern China Electric Power Design Institute Co.,Ltd.of Engineering Consulting Group,Wuhan 430071,China)
出处
《高电压技术》
EI
CAS
CSCD
北大核心
2021年第2期454-462,共9页
High Voltage Engineering
基金
国家自然科学基金(51777142
51907096)。
关键词
电力深度视觉
低漏报率
组合神经网络
缺陷识别
自适应融合
输电线路
electric power depth vision
low false negative rate
combined neural network
defect recognition
adaptive fusion
transmission line
