计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (21): 319-326.DOI: 10.3778/j.issn.1002-8331.2208-0163

• 工程与应用 • 上一篇    下一篇

基于旋转框定位的绝缘子污秽检测算法

王新良,纪昂志,李自强   

  1. 1.河南理工大学 物理与电子信息学院, 河南 焦作 454003
    2.许继电气股份有限公司,河南 许昌 461000
  • 出版日期:2023-11-01 发布日期:2023-11-01

Insulator Contamination Detection Algorithm Based on Rotating Frame Location

WANG Xinliang, JI Angzhi, LI Ziqiang   

  1. 1.School of Physics and Electronic Information Engineering, Henan Polytechnic University, Jiaozuo, Henan 454003, China
    2.XJ Electric Co., Ltd., Xuchang, Henan 461000, China
  • Online:2023-11-01 Published:2023-11-01

摘要: 为解决当前无人机巡检污秽绝缘子过程中受光照强弱影响大、背景复杂造成检测准确率低以及水平框并不能准确定位绝缘子等问题,提出一种改进R3Det的绝缘子污秽细粒度旋转目标检测算法。在特征提取部分,使用ConvNeXt做为主干特征提取网络,实现对绝缘子污秽细粒度特征的增强提取;同时采用PANet特征融合网络关联不同感受野特征。在检测头网络部分,使用对齐卷积和小尺度卷积,提升模型的检测性能以及增加分类的准确性;并利用Kullback-Leibler Divergence(KLD)优化损失函数,改善带有旋转角度信息的污秽绝缘子检测框的定位精确。实验结果表明,改进后的算法在自制绝缘子污秽数据集上的mAP可以达到90.6%,相较于原始网络提高了4.9个百分点,同时模型计算量降低了25.2%,能够准确有效地识别并定位出输电线路中的污秽绝缘子。

关键词: 无人机图像, 绝缘子污秽, 旋转目标检测

Abstract: An improved R3Det algorithm for fine-grained rotating target detection of contamination insulators is aiming at the problem that the current UAV inspection of fouled insulators is greatly influenced by the light intensity and low detection accuracy due to the complex background and the horizontal frame does not accurately locate the insulators is proposed. For the backbone of the network, ConvNeXt is used as the backbone feature extraction network to realize enhanced extraction of fine-grained features of insulator contamination. Meanwhile, the PANet is used to associate different receptive field features at the same time. In the part of detection head, aligned convolution and small-scale convolution are used to improve the detection performance of the model and increase the accuracy of classification. Kullback Leibler divergence(KLD) is also used to optimize the loss function to improve the positioning accuracy of the contaminated insulator detection frame with rotation angle information. The experimental results show that the mAP of the improved algorithm on the self-made insulator pollution data set can reach 90.6%, which is 4.9 percentage points higher than the original network, and the calculation amount of the model is reduced by 25.2%, which can accurately and effectively identify and locate the contaminated insulators in the transmission lines.

Key words: UAV image, insulator contamination, rotating object detection