基于改进的YOLOv3在电力设备红外检测中的应用  

The Application in Electric Power Equipment Infrared Object Detection Based on Improved YOLOv3

作　　者：陈寅 赵佰亭[1] CHEN Yin;ZHAO Baiting(School of Electrical and Information Engineering,Anhui University of Science and Technology,Huainan Anhui 232001,China)

机构地区：[1]安徽理工大学电气与信息工程学院,安徽淮南232001

出　　处：《兰州工业学院学报》2022年第5期29-34,共6页Journal of Lanzhou Institute of Technology

基　　金：安徽省自然科学基金(2108085ME158);安徽高校协同创新项目(GXXT-2020-54)。

摘　　要：针对变电站内电力设备的红外图像受到复杂背景、高遮挡、低对比度、相似目标特征的影响,而原始的YOLOv3算法模型很难对红外目标实现精准定位且模型过大无法部署到低计算能力的设备,提出一种改进的YOLOv3算法模型。引入MobileNetv3_Large主干网络替换原DarkNet53,以降低网络复杂度;在颈部网络添加空间金字塔池化(SPP)和DropBlock模块,以提升模型的泛化能力;加入K-means来优化整体的检测效果。结果表明:改进的YOLOv3超越原始算法,在测试数据集上的Map50达到了96.61%,检测速度达到了34.316帧/s。In response to the problems of complicated backgrounds,multiple occlusions,poor contrast and similar target features of infrared pictures of electric power equipment in substations,the original YOLOv3 algorithm model is difficult to achieve accurate localization of infrared targets as well as the model is too large to be deployed to low computing power devices,therefore an improved YOLOv3 algorithm model is proposed.To reduce the network complexity,the original DarkNet53 is replaced with the MobileNetv3_Large backbone network.The SPP layer and DropBlock module are added to the neck network to improve the generalization capability of the model.The K-means is introduced to improve the detection accuracy of the algorithm.The experimental results show that the improved YOLOv3 model surpasses the original algorithm,with a Map50 of 96.61% on the test dataset and a detection speed 34.316 frames/second,which is of practical value.

关 键 词：电力设备检测 YOLOv3 轻量级网络 红外图像 

分 类 号：TM726[电气工程—电力系统及自动化]