改进YOLOv8网络在绝缘子缺陷检测中的应用
Application of Improved YOLOv8 Network in Insulator Defect Detection
朱泓宇 1程换新 1骆晓玲2
作者信息
- 1. 青岛科技大学 自动化与电子工程学院,山东 青岛 266061
- 2. 青岛科技大学 机电工程学院,山东 青岛 266061
- 折叠
摘要
绝缘子缺陷目标小、分布零散等问题一直影响检测精度的提升,针对已有绝缘子缺陷检测方法的不足,基于注意力机制与多尺度融合方法,将 YOLOv8 网络增加了小目标检测层,并添加注意力与卷积混合模块(ACmix).提出了加权双向路径聚合网络(Bi-PANet)替代路径聚合网络(PANet),防止特征融合过程中原始特征的丢失,提高多尺度目标特征的融合度.使用 Wise-IOU作为回归损失函数,减少低质量标注的影响,加快网络收敛速度.对航拍图像中电力线路上的正常绝缘子和掉串绝缘子进行检测,结果表明提出的检测方法平均精度达到 93.2%,证明改进后的模型能够更好地识别绝缘子缺陷.
Abstract
The problem of small and scattered insulator defect targets has been affected the improvement of detection precision.In response to the shortcomings of existing insulator defect detection methods,the algorithm adds a small tar-get detection layer to the YOLOv8 network based on the attention mechanism and multi-scale fusion,and adds the self-at-tention and convolution(ACmix).A weighted bi-directional path aggregation network(Bi-PANet)is proposed instead of path aggregation network(PANet)to prevent the loss of original features during feature fusion and improve the fusion of multi-scale target features.Using Wise-IOU as the regression loss function,the influence of low-quality annotations is re-duced and the network convergence speed is accelerated.Experiments are conducted to detect normal insulators and dropped string insulators on power lines in aerial images,and the results show that the mean average accuracy of the proposed detection method reaches 93.2%,which proves that the improved model is able to better recognize insulator defects.
关键词
YOLOv8/ACmix注意力/Wise-IOU/绝缘子缺陷检测Key words
YOLOV8/ACmix attention/Wise-IOU/insulator defect detection引用本文复制引用
出版年
2024
NETL
NSTL
万方数据