College of Mechanical and Electrical Engineering, Chizhou University, Chizhou 247000, China.
College of Electronic and Information Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China.
Comput Intell Neurosci. 2022 Aug 17;2022:7113765. doi: 10.1155/2022/7113765. eCollection 2022.
Insulators identification and their missing defect detection are of paramount importance for the intelligent inspection of high-voltage transmission lines. As the backgrounds are complex, some insulators may be occluded, and the missing defect of the insulator is so small that it is not easily detected from aerial images with different backgrounds. To address the above issues, in this study, a cascaded You Only Look Once (YOLO) models are mainly explored to perform insulators and their defect detection in aerial images. Firstly, the datasets used for insulators location and missing defect detection are created. Secondly, a new model is proposed to locate the position of insulators, which is improved in the feature extraction network and multisacle prediction network based on previous YOLOv3-dense model. An improved YOLOv4-tiny model is used to conduct missing defect detection on the detected insulators. And then, the proposed YOLO models are trained and tested on the built datasets, respectively. Finally, the final models are cascaded for insulators identification and their missing defect detection. The average precision of missing defect detection can reach 98.4%, which is 5.2% higher than that of faster RCNN and 10.2% higher than that of SSD. The running time of the cascaded YOLO models for missing defect detection can reach 106 frames per second. Extensive experiments demonstrate that the proposed deep learning models achieve good performance in insulator identification and its missing defect detection from the inspection of high-voltage transmission lines.
绝缘子的识别及其缺失缺陷的检测对于高压输电线路的智能巡检至关重要。由于背景复杂，一些绝缘子可能会被遮挡，而且绝缘子的缺失缺陷很小，从不同背景的航拍图像中很难检测到。针对上述问题，本研究主要探索级联的 You Only Look Once（YOLO）模型在航拍图像中进行绝缘子及其缺陷检测。首先，创建了用于绝缘子定位和缺失缺陷检测的数据集。其次，提出了一种新的模型来定位绝缘子的位置，该模型在基于先前的 YOLOv3-dense 模型的特征提取网络和多尺度预测网络中得到了改进。使用改进的 YOLOv4-tiny 模型对检测到的绝缘子进行缺失缺陷检测。然后，分别在构建的数据集上对提出的 YOLO 模型进行训练和测试。最后，对最终模型进行级联，用于绝缘子识别及其缺失缺陷检测。缺失缺陷检测的平均精度可达 98.4%，比 faster RCNN 高 5.2%，比 SSD 高 10.2%。级联 YOLO 模型用于缺失缺陷检测的运行时间可达 106 帧/秒。大量实验表明，所提出的深度学习模型在从高压输电线路巡检中识别绝缘子及其缺失缺陷方面取得了良好的性能。
