基于改进YOLOv5s的输电线路红外图像多目标检测  

Multi-Target Detection of Infrared Images in Transmission Lines Based on Improved YOLOv5s

作　　者：杨春萍[1] 刘凯波 刘慕然 YANG Chunping;LIU Kaibo;LIU Muran(School of Electrical and Electronic Engineering,North China Electric Power University,Beijing 102206,China;International Education Institute,North China Electric Power University,Beijing 102206,China)

机构地区：[1]华北电力大学电气与电子工程学院,北京102206 [2]华北电力大学国际教育学院,北京102206

出　　处：《三峡大学学报(自然科学版)》2025年第3期105-112,共8页Journal of China Three Gorges University:Natural Sciences

基　　金：国网新疆电力有限公司科技项目(5230BD230003)。

摘　　要：为提高输电线路红外图像检测的可靠性和准确性,从实用化以及多目标检测角度出发,提出了一种基于改进YOLOv5s的输电线路检测及故障识别方法YOLOv5s-ECW.首先在Backbone部分增加跨空间学习的高效多尺度注意力机制,增强模型提取特征和多尺度融合的能力,减少了计算开销;其次在Neck部分引入上下文增强模块,减少信息冲突,提高了小目标与远距离目标的检测精度;最后将损失函数替换为Wise-IoU,使模型聚焦于普通质量的锚框,提升了检测效果.实验与测试结果表明,本文提出的方法YOLOv5s-ECW与原YOLOv5s相比,平均精度、精确率、召回率分别提升了3.9%、4.0%、4.5%,并且针对5种电力设备以及可能故障点的检测能力得到不同程度的增强,方法更加实用化.To improve the reliability and accuracy of infrared image detection in transmission lines,a method YOLOv5s ECW based on the improved YOLOv5s for the detection and fault recognition in transmission line is proposed in this paper from the perspectives of practicality and multi-target detection.Firstly,an efficient multi-scale attention mechanism for cross spatial learning is added to the Backbone section to enhance the model's ability in order to extract the features and perform the multi-scale fusion.Furthermore,the computational overhead is reduced;Secondly,a context enhancement module is introduced in the Neck section to reduce the information conflicts and improve the detection accuracy of small and distant targets;Finally,the loss function is replaced with Wise IoU,which focused the model on anchor boxes of ordinary quality and improved the detection performance.The experiments and tests show that the proposed method YOLOv5s-ECW has improved the average accuracy by 3.9%,the accuracy by 4.0%,the recall by 4.5%compared with that of the original YOLOv5s.Moreover,its detection ability for five types of power equipment and possible fault points has been enhanced in various degrees.The method is more practical.

关 键 词：YOLOv5s 红外图像 多尺度融合 小目标检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]