基于改进的YOLOv5绝缘子检测与识别算法

Detection and Recognition Algorithm for Power Insulator in Complex Background Environment

摘要:

针对复杂背景环境下高压电力绝缘子日常维护和缺陷检测等智能化巡检的不足,提出一种基于改进YOLOv5的电力绝缘子检测与识别算法。首先,针对经典YOLOv5算法对于相似绝缘子误检率较高问题,通过添加通道模块和空间注意力并联模块,增强绝缘子目标的特征和位置信息以降低误检率；接着,针对多检缺陷目标的问题提出基于面积比的抑制算法,利用基于损失函数和后处理的改进措施进一步筛选缺陷预测框；最后,对于复杂背景遮挡绝缘子导致漏检问题,分别在检测器中采用CIoU Loss回归损失。实验测试表明,所提出算法不仅解决相似绝缘子的误检和遮挡漏检问题,而且还提高模型的精度和速度,其精度mAP和推理速度分别为0.886和65.2FPS,相比经典YOLOv5算法分别提高11.4和5.8FPS。

Abstract:

Aiming at the shortage of intelligent inspection such as daily maintenance and defect detection of high-voltage power insulators in complex background environment, an algorithm for detection and identification of power insulators in complex background environment is proposed. Firstly, this paper addresses the problem of high false detection rate of similar insulators in YOLOv5 algorithm by adding a channel module to improve the focus of the algorithm on the target features and adding a spatial attention parallel module to obtain insulator location information; then, to address the issue of multiple defect targets, a suppression algorithm based on area ratio is proposed , and by using improvement measures based on loss function and post-processing further the screen defect prediction boxes is used also. Finally, for the problem of missing insulators due to complex background occlusion, CIoU Loss regression loss is used in the detector. Experimental tests show that the proposed algorithm not only solves the problems of false detection of similar objects and missed detection of occluded insulators, but also improves the accuracy and speed of the model, with the accuracy mAP and inference speed of 0.886 and 65.2 FPS, respectively, which are 11.4 and 5.8 FPS higher than the traditional YOLOv5 classical algorithm.