基于改进YOLOv8n的电力输电线路鸟巢和绝缘子检测

Detection of Bird"s Nest and Insulator in Power Transmission Lines Based on the Improved YOLOv8n

摘要:

YOLOv8n泛化性能和对部分小目标检测效果欠佳，且训练时参数量较大，为了更好地检测出输电线路上的鸟巢和绝缘子，采用一种动态标签分配策略DynATSS，在每次迭代中引入预测，更好地定义了正负样本，能选择更多的经预测为高质量的样本作为正样本，预测相较于预定义Anchor也更加准确；将原始检测头替换为新检测头ODH，提高了模型的检测精度，同时参数量也得到了相应的减少，原始耦合头引入的分类与回归任务之间的冲突也得到了有效解决；将模型的原始损失函数C-IoU替换为W-IoU，在锚框与目标重合度较高时削弱了几何因素的惩罚，同时较少的干预训练使得模型的泛化能力也得到了提高。采用5432幅图像进行了训练，结果表明，改进后的YOLOv8nDOW算法，mAP50较原始的YOLOv8n模型提升了1.7%，mAP50-95提升了1.5%，符合巡检输电线路的准确性、实时性要求。

Abstract:

The generalization performance and detection effect of YOLOv8n on some small targets are not good, and the number of parameters is large during training. In order to better detect the nests and insulators on the transmission line, a dynamic label allocation strategy DynATSS is used to introduce prediction in each iteration to better define the positive and negative samples, and more samples that are predicted to be high-quality can be selected as positive samples, and the prediction is more accurate than the predefined anchor. The original detection head was replaced with the new detection head ODH, which improved the detection accuracy of the model, and the number of parameters was reduced accordingly, and the conflict between the classification and regression tasks introduced by the original coupling head was also effectively resolved. The original loss function C-IoU of the model is replaced by W-IoU, which weakens the punishment of geometric factors when the anchor frame coincides with the target with a high degree of coincidence, and the generalization ability of the model is also improved with less intervention training. The results show that the improved YOLOv8nDOW algorithm improves mAP50 by 1.7% and mAP50-95 by 1.5% compared with the original YOLOv8n model, which meets the accuracy and real-time requirements of inspection of transmission lines.