计算机与现代化 ›› 2023, Vol. 0 ›› Issue (02): 78-82.

• 人工智能 • 上一篇    下一篇

基于改进YOLOv5的输电线路多目标检测

  

  1. (1. 南京工程学院电力工程学院,江苏 南京 211167; 2. 南京工程学院计算机工程学院,江苏 南京 211167)
  • 出版日期:2023-04-10 发布日期:2023-04-10
  • 作者简介:汤浩威(1996—),男,江苏宿迁人,硕士研究生,研究方向:图像处理,电力设备故障检测,E-mail: 1925774316@qq.com; 姚军财(1979—),男,教授,博士,研究方向:图像和视频处理,目标检测和计算机视觉与模式识别,E-mail: yaojcnj@njit.edu.cn; 姚聪颖(1998—),男,江苏南京人,硕士研究生,研究方向:电力设备故障检测; 孙颖(1998—),女,江苏扬州人,硕士研究生,研究方向:机器学习; 裴星懿(1999—),男,江苏南京人,硕士研究生,研究方向:电力设备故障检测; 宋春晓(1997—),女,山东德州人,硕士研究生,研究方向:信息安全。
  • 基金资助:
    国家自然科学基金资助项目(61301237);江苏省自然科学基金资助面上项目(BK20201468);南京工程学院高层次引进人才基金资助项目(YKJ201981)

Multi-target Detection of Transmission Lines Based on Improved YOLOv5

  1. (1.School of Electrical Engineering, Nanjing Institute of Technology, Nanjing 211167, China;
    2.School of Computer Engineering, Nanjing Institute of Technology, Nanjing 211167, China)
  • Online:2023-04-10 Published:2023-04-10

摘要: 针对当前目标检测网络层数加深、参数量和计算量加大,造成实时性差等问题,为了实现对输电线路部件的识别与检测,提出一种基于改进YOLOv5的输电线路多目标检测算法。首先,使用ShuffleNetv2结构作为网络特征提取的主干结构,减少网络的参数量;然后,将PANet网络中的BottleneckCSP 改为Light_CSP模块,加快特征融合的速度;其次,使用CIoU loss、DIoU-NMS方法减少预测框的位置损失和漏检问题。最后,为了验证所提算法的有效性,利用输电线路图像数据集进行训练与测试。结果表明,改进YOLOv5的参数量为7.5×106,浮点计算量为10.9,平均精度达到了87.5%,FPS达到69.2,能够满足输电线路部件检测的精度、轻量化与实时性要求。

关键词: 智能巡检, 目标检测, YOLOv5, 输电线路

Abstract: In order to realize the identification and detection of transmission line components, a multi-target detection algorithm for transmission lines based on improved YOLOv5 is proposed for the current problems of deepening the number of target detection network layers, increasing the number of parameters and computation, resulting in poor real-time performance. Firstly, the number of parameters in the network was reduced by using the shuffleNetv2 structure as the backbone structure for network feature extraction. Secondly, the BottleneckCSP in the PANet network is changed to a Light_CSP module to speed up feature fusion. Thirdly, the CIoU loss, DIoU-NMS method is used to reduce the loss of position of the prediction frame and the problem of missed detection. Finally, in order to verify the effectiveness of the proposed algorithm, a transmission line image dataset was used for training and testing The results show that the improved YOLOv5 has a parametric count of 7.5×106, a floating point computation of 10.9, an average accuracy of 87.5% and an FPS of 69.2, which meets the requirements for accuracy, lightness and real-time inspection of transmission line components.

Key words: intelligent inspection, target detection, YOLOv5, transmission line