摘要
针对绝缘子所处环境的复杂性及掉串缺陷在航拍图像中占比较小的问题,提出了一种基于改进YOLOv5算法的输电线路绝缘子掉串缺陷识别方法。以YOLOv5算法为绝缘子检测的基础,在SPP模块将原网络的池化结构改为SoftPool;引入DIoU_NMS代替NMS,并将损失函数设置为CIoU+DIoU_NMS,最后以PSO融合K-means算法优化初始锚框并改变最终输出通道个数。通过实验结果表明,改进后的YOLOv5对绝缘子及掉串缺陷双目标检测的平均检测精度比原始YOLOv5提高3.7%,上述方法能较为有效地、准确地识别绝缘子掉串缺陷。
In view of the complexity of insulator environment and the small proportion of drop defects in aerial im-ages,an improved YOLOv5 algorithm is proposed to identify insulator drop defects in transmission lines.Firstly,YOLOv5 algorithm was used as the basis of insulator detection,and the original network pooling structure was changed to Soft Pool in SPP module.DIo U_NMS was introduced to replace NMS,and the loss function was set as CI-o U+DIo U_NMS.Finally,PSO fusion K-means was used to optimize the initial anchor frame and change the final number of output channels.Experimental results show that the average detection accuracy of the improved YOLOv5 is 3.7%higher than that of the original YOLOv5,and the method can identify insulator drop defects more effectively and accurately.
作者
乔钰彬
范菁
张宜
肖云波
QIAO Yu-bin;FAN Jing;ZHANG Yi;XIAO Yun-bo(College of Electrical and Information Engineering,Yunnan Minzu University,Kunming Yunnan 650531,China)
出处
《计算机仿真》
北大核心
2023年第7期132-137,共6页
Computer Simulation
基金
国家自然科学基金项目(61540063)
云南省应用技术研究计划项目(2018FD055)。
关键词
绝缘子
深度学习
目标检测
Insulator
Deep learning
Target detection
