基于改进YOLOv5的变电站表计缺陷检测算法  

Substation meter defect detection algorithm based on improved YOLOv5 network

作　　者：鲍文霞[1] 袁牧 梁栋 王年[1] 杜翔 BAO Wenxia;YUAN Mu;LIANG Dong;WANG Nian;DU Xiang(School of Electronic Information Engineering,Anhui University,Hefei 230601,China;College of Internet,Anbui University,Hefei 230039,China)

机构地区：[1]安徽大学电子信息工程学院,安徽合肥230601 [2]安徽大学互联网学院,安徽合肥230039

出　　处：《安徽大学学报（自然科学版）》2024年第1期50-56,共7页Journal of Anhui University(Natural Science Edition)

基　　金：国家重点研发计划项目(2020YFF0303803)。

摘　　要：准确检测变电站中的设备缺陷并及时进行处理是保证电力系统安全运行的重要措施.针对表计缺陷图像背景复杂、目标尺寸不一、外形差别大等问题,提出基于改进YOLOv5(you only look once的第5个版本)的变电站表计缺陷检测算法.为了提高泛化能力、解决训练过程中样本不平衡问题,利用旋转和改变图像亮度的方法进行数据增广.通过引入坐标注意力机制,在聚焦缺陷特征的同时,能突出缺陷特征的差异.为了使边界框回归更快速准确,将EDIOU loss(effective distance intersection over union loss)代替CIOU loos(complete intersection over union loss).实验结果表明:6种算法中,该文算法的准确度、召回率和mAP(mean average preciscion)均最高,分别达85.1%,86.6%,87.3%.因此,该文算法具有优越性.Accurate detection and timely replacement of meters in substations are important measures to ensure the safe operation of power systems.An improved YOLOv5(the fifth version of you only look once)based meter defect detection algorithm for substation was proposed to solve the problems of meter defect image,such as complex background,different target size and large appearance difference.In order to improve the generalization ability and solve the problem of sample imbalance in the training process,the rotation and changing the brightness of the image were used to increase the data.The coordinate attention(CA)mechanism was introduced to highlight the differences between defect features while focusing on defect features.To make bounding box regression faster and more accurate,EDIOU loss(effective distance intersection over union loss)was replaced by CIOU loos(complete intersection over union loss).The experimental results showed that the P(prcision),R(recall)and mAP(mean average precision)of algorithm of this paper were the highest among the six algorithms,reaching 85.1%,86.6%and 87.3%,respectively.Therefore,the algorithm of this paper had superiority.

关 键 词：表计缺陷 YOLOv5 数据增广 注意力机制 损失函数 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]