首页|基于改进YOLOv5算法的接触网绝缘子定位方法

基于改进YOLOv5算法的接触网绝缘子定位方法

扫码查看
[目的]针对高速铁路接触网绝缘子在复杂背景下检测效率不高的问题提出一种检测算法.[方法]首先对样本数据集进行大规模扩充,在原有YOLOv5s算法的基础上,为有效的提升模型的表征力,增加ECA注意力机制,进行无降维的跨信道方式来聚焦绝缘子位置信息;使用BiFPN特征金字塔网络,进行多尺度的特征融合来丰富语义信息;选用Meta-ACON自适应控制激活函数,在函数允许的最大范围内,严格把控函数的上下限,防止模型出现失控现象;将原有GIOU损失函数更换为EIOU损失函数,从梯度的角度对锚框进行更深一步的划分,进而提升网络的收敛速度.[结果]实验结果表明,通过对YOLOv5s改进后的检测算法,可以对绝缘子进行更精确的定位与识别,准确率达到了99.4%.[结论]所提出的检测算法为绝缘子定位检测提供了更加准确快捷的方法.
Catenary Insulator Positioning Method Based on Improved YOLOv5 Algorithm
[Purpose]A detection algorithm is proposed to address the issue of low detection efficiency of high-speed railway contact line insulators in complex backgrounds.[Methods]On the basis of the original YOLOv5s algorithm,in order to effectively improve the representation power of the model and increase the ECA attention mechanism,a cross-channel method without dimensionality reduction is carried out to focus on the position infor-mation of insulators.The BiFPN feature pyramid network is used to enrich the semantic information by multi-scale feature fusion.The Meta-ACON adaptive control activation function is selected and the upper and lower limits of the function is strictly controlled within the maximum range allowed by the function to prevent the mod-el from running out of control.The original GIOU loss function is replaced with the EIOU loss function,and the anchor box is further divided from the perspective of gradient,so as to improve the convergence speed of the net-work.[Results]Acoording to the experimental results,the improved detection algorithm of YOLOv5s can be used to locate and identify the insulator more accurately,and the accuracy rate reaches 99.4%.[Conclusion]The proposed detection algorithm provides a more accurate and faster method for insulator positioning detection.

insulatorsYOLOv5sstyle migrationattention mechanismbidirectional fusion feature network

刘仕兵、周诗涵、但业光

展开 >

华东交通大学电气工程学院,江西南昌 330013

绝缘子 风格迁移 YOLOv5s 注意力机制 双向融合特征网络

轨道交通基础设施性能监测与保障国家重点实验室开放课题

GJJ210652

2024

华东交通大学学报
华东交通大学

华东交通大学学报

CSTPCD
影响因子:0.748
ISSN:1005-0523
年,卷(期):2024.41(1)
  • 16