基于YOLOv5-nS算法的绝缘子串销钉检测方法

Insulator string pin detection method based on YOLOv5-nS algorithm

针对变电站绝缘子更换机器人在单片绝缘子更换中多姿态的绝缘子销钉检测困难和检测速度慢的问题,提出了基于YOLOv5所改进的一种快速、可靠的绝缘子销钉检测方法YOLOv5-nS。首先,针对原网络Backbone部分提出了一种新的骨干网络结构,该结构基于ShuffleNetV2所改进的nSNet所构建;其次,对网络的Neck进行了轻量化的同时增加特征融合层;再次,改变边界框回归损失函数,以提升边界框预测的准确率;最后,进行了先验框(Anchor)尺度重选取。实验结果表明:改进后的YOLOv5-nS算法模型在精度基本不变的情况下,参数减少了89.7%,模型尺寸减小了87.5%,帧率提升了7 f/s。

When a substation insulator replacement robot is used to replace single-piece insulator,it is hard to detect the multi-gesture insulator pins and the detection speed is slow.To solve these defects,a fast and reliable insulator pin detection method YOLOv5-nS is proposed on the basis of YOLOv5.First,a new backbone network structure is proposed for the Backbone part of the original network.Backbone is built on the basis of nSNet obtained by improving ShuffleNetV2;Secondly,a feature fusion layer is added at the same time of lightening the Neck of network;And further,by changing the bounding box regression loss function,the prediction accuracy of bounding box is lifted;Finally,Anchor scale is reselected.The experimental results show that,under the condition that the accuracy of the improved YOLOv5-nS algorithm model is kept basically unchanged,the parameters are reduced by 89.7%,the model size is reduced by 87.5%,and the frame rate is increased by 7 f/s.