摘要
针对目前绝缘子缺陷检测方法速度较低、网络复杂度高且小目标缺陷难以准确检测的问题,提出一种轻量化绝缘子缺陷检测模型——P-YOLOv7-tiny。对主干网络的高效层聚合网络(Efficient Layer Aggregation Network, ELAN)模块进行轻量化处理,设计出P-ELAN模块以减少模型参数量,提高模型检测速度;将坐标注意力(Coordinate Attention, CA)机制与原模型结构CSPSPP进行融合,设计出CA-CSPSPPS模块使模型更加关注绝缘子缺陷特征,提高对缺陷的检测准确率;采用定位损失函数(WIoUv3 Loss)计算损失,将较小的梯度增益分配给低质量的锚框以减少有害梯度,提高模型的定位性能。实验结果表明,P-YOLOv7-tiny可以快速准确地检测缺陷,其中mAP@0.5达到了98.3%,召回率达到了95.3%,模型参数量为3.1 M,计算量为7.0 GFLOPs。相较原模型YOLOv7-tiny,所提模型适合部署到边缘设备对绝缘子缺陷进行实时检测。
In view of problems of low detection speed,high network complexity,and difficulty in accurately detecting small target defects in current insulator defect detection methods,a lightweight insulator defect detection model called P-YOLOv7-tiny is proposed.Firstly,lightweight processing is made on the Efficient Layer Aggregation Network(ELAN)module of the backbone network,and the P-ELAN module is designed to reduce the model parameters and improve the detection speed.Secondly,the Coordinate Attention(CA)mechanism is fused with CSPSPP to design the CA-CSPSPPS module,which allows the model to focus more on insulator defect features and improve the detection accuracy of defects.Finally,the localization loss function(WIoUv3 Loss)is used to calculate the loss,allocating smaller gradient gains to low-quality anchor boxes to reduce harmful gradients and improve the model s localization performance.Experimental results show that P-YOLOv7-tiny can quickly and accurately detect defects,with an mAP@0.5 of 98.3%and a recall rate of 95.3%.The model has 3.1 M parameters and a computational cost of 7.0 GFLOPs.Compared to the original YOLOv7-tiny model,the model parameters are reduced by 48.3%,the computational cost is reduced by 46.9%,and the recall rate is improved by 1.2%.The proposed model is suitable for deploying to edge equipment to detect insulator defects in real time.
作者
刘修政
王波
LIU Xiuzheng;WANG Bo(School of Mechanical Engineering,Anhui University of Science and Technology,Huainan 232001,China;School of Mechanical and Electrical Engineering,Chuzhou University,Chuzhou 239000,China)
出处
《无线电工程》
2024年第10期2305-2314,共10页
Radio Engineering
基金
安徽省教育自然科学基金(KJ2021A1086)
安徽省高校优秀拔尖人才培育项目(gxgnfx2022071)。