作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2024, Vol. 50 ›› Issue (5): 200-208. doi: 10.19678/j.issn.1000-3428.0067812

• 图形图像处理 • 上一篇    下一篇

高铁接触网绝缘子检测算法研究

刘仕兵, 周诗涵   

  1. 华东交通大学电气工程学院, 江西 南昌 330013
  • 收稿日期:2023-06-07 修回日期:2023-07-25 发布日期:2023-09-05
  • 通讯作者: 刘仕兵,E-mail:liucyier@163.com E-mail:liucyier@163.com
  • 基金资助:
    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)。

Research on Insulator Detection Algorithm for High-Speed Rail Contact Network

LIU Shibing, ZHOU Shihan   

  1. School of Electrical Engineering, East China Jiaotong University, Nanchang 330013, Jiangxi, China
  • Received:2023-06-07 Revised:2023-07-25 Published:2023-09-05
  • Contact: 刘仕兵,E-mail:liucyier@163.com E-mail:liucyier@163.com

摘要: 针对已有高速铁路接触网绝缘子目标检测算法通常存在检测精度不高且忽视了绝缘子方向的问题,为了能更好地满足智能化巡检需求,提出一种基于改进YOLOv5的绝缘子旋转目标检测算法。首先,引入协调注意力(CA)和十字交叉注意力机制,高效提取绝缘子的有效特征及位置信息,同时利用骨干网络RepVGG架构,有效提升模型表征力和检测速度;在检测头的骨干网络中,采用对齐卷积(AC)模块解决了绝缘子目标的倾斜和特征不对齐问题,进一步调整了预测框与实际目标的对齐程度;最后,采用旋转完全交并比(R-CIoU)计算旋转损失函数,可以更好地实现对预测框的精准定位。实验结果表明,该算法可以实现对绝缘子不同方向的检测,在提升检测速度的同时平均精度均值(mAP)达到97.5%,能更好地满足绝缘子目标检测的需求。

关键词: 旋转目标检测, YOLOv5网络结构, 绝缘子, 对齐卷积, 注意力机制

Abstract: To satisfy the requirements of intelligent inspection at a greater level, a rotated insulator target detection algorithm based on the improved YOLOv5 is proposed to solve the inadequacy of conventional detection algorithms used for contact networks in high-speed rails, e.g., low accuracy and non-consideration of insulation direction. First, Coordinated Attention (CA) and criss-cross attention mechanisms are introduced to efficiently extract the effective features and position information of insulators. The Reparameterization Visual Geometry Group (RepVGG) backbone network architecture is used to effectively improve the model representation and detection speed. In the backbone network of the detection head, the Alignment Convolution (AC) module is used to solve the tilt and feature misalignment of the insulator target, as well as to adjust the alignment degree of the prediction frame to the actual target. Finally, the Rotation Complete Intersection over Union (R-CIoU) is used to calculate the rotation loss function, which can be used to accurately position the prediction frame. Experimental results show that the proposed algorithm can detect different directions of insulators and that the mean Average Precision (mAP) can reach 97.5% while the detection speed is improved, thus satisfying the requirements of insulator target detection at a greater level.

Key words: rotated target detection, YOLOv5 network structure, insulator, Align Convolution(AC), attention mechanism

中图分类号: