基于改进YOLO的高压配电网污闪绝缘子无人机检测方法

邱略能; 郑志祥; 孟德威

基于改进YOLO的高压配电网污闪绝缘子无人机检测方法

引用本文：邱略能，郑志祥，孟德威.基于改进YOLO的高压配电网污闪绝缘子无人机检测方法[J].计算技术与自动化,2025,(3):123-127

摘要点击次数: 358

全文下载次数: 219

作者	单位
邱略能，郑志祥，孟德威	（国网衢州供电公司,浙江衢州 324000）

中文摘要:针对目前无人机下高压配电网绝缘子检测存在数据收集过程复杂、预测精度低的问题，提出了一种基于图像处理技术的高压配电网污闪绝缘子无人机检测方法。引入CA机制，通过对通道注意力进行分解，将位置信息嵌入到通道注意力中，从而增强网络在复杂背景下检测绝缘子的能力。使用Bi-FPN代替原始的PANet特征融合框架，通过对每个尺度进行加权以平衡不同尺度上的特征信息，从而增强了网络检测小目标的能力。实验阶段，通过消融实验，与原始网络相比，所提改进YOLO模型的mAP提高约2.7%，Recall提高约4.0%。与SSD、RetinaNet、YOLOv4、YOLOv5x和YOLOv7相比，所提模型mAP最高，约为90.3%。实验结果验证了所提方法的有效性和实用性，该模型具有广阔的应用前景。

中文关键词:高压配电网电力巡检深度学习特征提取注意力机制多尺度特征

A UAV Detection Method for Pollution Flashover Insulators in High Voltage Distribution Networks Based on Improved YOLO

Abstract:Aiming at the problems of complex data collection process and low prediction accuracy in the detection of high-voltage distribution network insulators under unmanned aerial vehicles (UAVs), a UAV detection method for high-voltage distribution network pollution flashover insulators based on image processing technology is proposed. Introducing CA mechanism to decompose channel attention and embed position information into channel attention, thereby enhancing the network’s ability to detect insulators in complex backgrounds. The use of Bi-FPN instead of the original PANet feature fusion framework enhances the network’s ability to detect small targets by weighting each scale to balance feature information at different scales. In the experimental stage, through ablation experiments, compared with the original network, the mAP of the proposed improved YOLO model increased by about 2.7%, and the Recall increased by about 4.0%. Compared with SSD, RetinaNet, YOLOv4, YOLOv5x, and YOLOv7, the proposed model has the highest mAP at approximately 90.3%. The experimental results validate the effectiveness and practicality of the proposed method, and the model has broad application prospects.

keywords:high voltage distribution network electric power inspection deep learning feature extraction attention mechanism multiscale features

查看全文 查看/发表评论 下载pdf阅读器