基于改进YOLOv5的绝缘子缺陷检测算法

绝缘子缺陷检测是电网巡检过程中重要的一环，为提高绝缘子缺陷检测的精度，该文提出一种基于改进YOLOv5算法的绝缘子缺陷检测算法——YOLOv5t，能够在保证网络运行速度的条件下，提升网络的检测精度.该算法在YOLOv5s的基础上，将三重注意力机制(triplet attention)添加到骨干网络中，给予每个特征通道不同的权重，以提高网络的检测精度；并采用CIoU Loss作为网络回归损失的损失函数，提升网络的收敛速度；同时将Soft-NMS作为网络的预测结果处理方法，降低网络的漏检率.YOLOv5t与几种常用的缺陷检测网络的对比实验结果表明，YOLOv5t的准确率达到97.2%，召回率达到98%，平均精度均值达到99.1%，较YOLOv5s算法分别提升了0.9%、5.1%和2.1%，并且检测速度没有受到影响.

Insulator defect detection algorithm based on improved YOLOv5

Insulator defect detection plays an important part in the inspection process of the power system. To improve the accuracy of insulator defect detection， a new detection method， YOLOv5t， was proposed based on the YOLOv5 network. The YOLOv5t method was able to increase the accuracy while assuring the speed of network detection. Based on the YOLOv5s network model， theTtriplet Attention was added to the backbone network of the model， and different feature channels were given different weights to improve the detection accuracy of the network. CIoU loss function is used to calculate the regression loss of the network to improve the convergence speed of network. At the same time， soft-NMS is used as the network prediction result processing method to reduce the missed detection rate. Experimental results show that the accuracy， recall and mean accuracy of YOLOv5t in insulator defect detection are 97.2%， 98% and 99.1%， respectively. In comparison to YOLOv5s， YOLOv5t had better accuracy， recall and mean accuracy， which were increased by 0.9%， 5.1% and 2.1%， respectively. And the detection speed is not affected.