摘要
针对架空地线破损检测场景,传统计算机视觉的手工特征泛化性差,主流通用目标检测网络迁移效果差等问题,提出一种以YOLOv5n作为主干网络的目标检测算法。首先,提出一种交叉上下文特征提取模块替换原主干特征提取模块,提升了模型针对遥感场景的特征提取能力。其次,颈部部分引入DAnet中的位置注意模块和通道注意模块提升模型复杂背景下多尺度目标的检测能力,抑制显著无关目标的特征,提升模型检测精度,并且进行了跨层特征图连接和增添了一层小目标检测层,提升模型对于细微小目标的感知能力。最后设计了一种类内聚合且类间分离的损失函数,显示引导模型提取更具可判别性的特征。相较于原YOLOv5n结构,改进算法在精确率方面提升了9.1%,召回率方面提升了7.4%,mAP50提升了10.7%。实验结果表明:改进YOLOv5算法能够高效且准确地识别出复杂背景下的架空破损地线目标。
In aerial ground wire damage detection,the generalization of traditional computer vision’s manual features is poor,and the migration effect of mainstream general target detection network is bad.To address these issues,a target detection algorithm using YOLOv5n as the backbone network was proposed.First,a cross context feature extraction module was proposed to replace the original backbone feature extraction module,which improved the feature extraction ability of the model for remote sensing scenes.Second,the position attention module and channel attention module in DAnet were introduced into the neck part to improve the detection ability of multi-scale targets in the complex background of the model,suppress the features of significantly unrelated targets,and improve the detection accuracy of the model.In addition,cross layer feature maps were connected and a layer of small target detection was added to improve the perception ability of the model for small and subtle targets.Finally,a loss function was designed for intra-class aggregation and inter-class separation,to guide the model to extract more discriminative features.Compared with the original YOLOv5n structure,the improved algorithm has improved the accuracy rate by 9.1%,the recall rate by 7.4%,and the mAP50 by 10.7%.The experimental results showed that the improved YOLOv5 algorithm can effectively and accurately identify the aerial broken ground wire targets in a complex background.
作者
雷泽宇
杨暘
杨雄
杨生兰
卢锐
马晓虎
LEI Zeyu;YANG Yang;YANG Xiong;YANG Shenglan;LU Rui;MA Xiaohu(State Grid Sichuan Extra High Voltage Company,Chengdu 610000,China)
出处
《国外电子测量技术》
2025年第2期175-184,共10页
Foreign Electronic Measurement Technology
基金
国家电网四川省电力公司科技项目(521962230001)。
