基于YOLOv5的配电网绝缘子缺陷分级检测

Fault Hierarchical Detection of Insulator in Distribution Network Based on YOLOv5 Neural Network

针对目前配电网绝缘子巡检因图像背景复杂、检测目标较小、缺陷形态多样、阴影遮挡导致误检漏检等原因而造成的检测准确率低的问题,提出一种用于配电线路绝缘子缺陷检测的先定位后分类的分级检测策略。首先采用YOLOv5网络定位绝缘子区域,在此基础上通过DenseNet网络进一步区分绝缘子区域是否存在故障。通过第二级DenseNet201网络能够识别出在遮挡情况下特征表达能力不足的故障绝缘子,同时排除因背景杂物造成的误检。实验结果表明,所提改进YOLOv5s绝缘子定位网络的检测时间为平均每张图片11.8 ms,检测均值平均精度达到97.4%,均优于原有YOLOv5模型。

Aiming at the low detection accuracy of current distribution network insulator inspection due to complex image backgrounds,small detection targets,diverse defect shapes,and false detection and missed detection caused by shadow occlusion,a hierarchical detection strategy for distribution line insulators fault detection is proposed which is based on localization first and then classification.First,the YOLOv5 network is used to locate the insulator area,and on this basis,the DenseNet is used to further distinguish whether there is a fault in the insulator area.Through the second-level DenseNet201 network,fault insulators with insufficient feature expression ability under occlusion can be identified,and false detection caused by background debris can be eliminated at the same time.The experimental results show that the detection time of the proposed improved YOLOv5s insulator location network is 11.8 ms per image on average,and the average detection accuracy reaches 97.4%,which are better than the original YOLOv5 model.