摘要
针对传统输电线路绝缘子串破损情况存在识别准确率低,小目标检测困难的现象,文中提出一种改进的YOLOv5s绝缘子串破损检测算法。首先,将Swin Transformer编码结构嵌入到原始C3模块中,同时将Swin Transformer滑动窗口中窗口的Patch数量设计为8×8大小,使其能够提取到更多的特征信息,该模块能够加强深层的语义特征信息,提升对缺陷目标的检测能力;其次,设计了一种C3CBAM结构代替neck层的C3模块,同时将标准卷积变为深度可分离卷积,使其在降低参数量的同时提高了检测精度;最后,在特征融合网络部分使用BiFPN,提升模型的特征融合能力,改善对绝缘子串破损目标的检测能力。实验结果表明,改进后的YOLOv5s模型与原模型相比,检测的平均精度mAP达到92%,模型参数为6.67M,模型大小为13.9MB,检测速度为208FPS,证明了改进后方法的有效性。
Aiming at the phenomenon of low recognition accuracy and difficulty in detecting small targets in the case of insulator string breakage of traditional transmission lines,the author proposes an improved YOLOv5s insulator string breakage detection algorithm.Firstly,the Swin Transformer coding structure is embedded into the original C3 module,and at the same time,the number of Patch in the window in the Swin Transformer sliding window is designed to be 8×8 in size,so that more feature information can be extracted,and this module can strengthen the deep semantic feature information and improve the detection ability of defective targets;secondly,a C3CBAM structure instead of the C3 module of the neck layer,and at the same time,the standard convolution is changed into the depth separable convolution,so that it can improve the detection accuracy while reducing the number of parameters;finally,BiFPN is used in the part of the feature fusion network,which enhances the feature fusion capability of the model and improves the detection capability of the insulator string breakage target.The experimental results show that the improved YOLOv5s model,compared with the original model,achieves an average detection accuracy mAP of 92%,a model parameter of 6.67 M,a model size of 13.9 MB,and a detection speed of 208 FPS,which proves the effectiveness of the improved method.
作者
王素珍
邓成禹
葛润东
吕基岳
WANG Suzhen;DENG Chengyu;GE Rundong;LU Jiyue(School of Information and Control Engineering,Qingdao University of Technology,Qingdao 266525,China)
出处
《电瓷避雷器》
2025年第4期109-118,共10页
Insulators and Surge Arresters
基金
山东省自然科学基金项目(编号:ZR2020QF101,ZR2021MF024)。
作者简介
王素珍(1975-),女,博士,副教授,主要研究方向为数字孪生与人工智能。E-mail:417322899@qq.com;通信作者:邓成禹(1995-),男,硕士,主要研究方向为人工智能与图像处理。E-mail:476355900@qq.com;葛润东(1999-),硕士,主要研究方向为数据挖掘。E-mail:2061130584@qq.com;吕基岳(1991-),硕士,主要研究方向为人工智能。E-mail:87928916@qq.com。
