高压电器2025，Vol.61Issue(6)：102-112,11.DOI:10.13296/j.1001-1609.hva.2025.06.012
基于改进YOLOv7的多尺度特征提取绝缘子缺陷检测算法
Insulator Defect Detection Algorithm for Multi-scale Feature Extraction Based on Improved YOLOv7
摘要
Abstract
In view of such issues as small defect sizes in insulation images,abundant background interference factors and imbalanced difficulty and easy sample of insulator,a multi-scale fusion YOLOv7(MSF-YOLOv7)algorithm is proposed and applied to the defect detection of insulator of power transmission lines.Firstly,the multi-scale convolu-tion block attention module(MS-CBAM)is introduced to deeply aggregate feature maps with rich semantic informa-tion,enhancing the network's detection performance for target at different scales.Then,a newly developed global spa-tial pyramid pooling-fast(GSPPF)module is used in the backbone network to incorporate global background informa-tion,mitigating the influence of complex backgrounds.In view of imbalanced sample distribution,the Focaler-CIoU loss function is introduced to focus on different defect targets,speeding up the model's convergence rate.The experi-mental results indicate that the mAP50 of MSF-YOLOv7 model proposed in this paper is up to 88.1%,the precision and recall rates are up to 90.3%and 83.1%,which are 6.3%,7.9%and 6.3%higher than those of the YOLOv7 algo-rithm and,at the same time,the number of parameters and GFLOPs are reduced by 13.38%and 2.95%,respectively.关键词
YOLOv7/绝缘子缺陷检测/多尺度特征提取/注意力机制/空间金字塔池化Key words
YOLOv7/defect detection of insulator/muti-scale feature extraction/attention mechanism/spatial pyramid pooling引用本文复制引用
孙虹,郭乂菡,雷经发,赵汝海,李永玲,张淼..基于改进YOLOv7的多尺度特征提取绝缘子缺陷检测算法[J].高压电器,2025,61(6):102-112,11.基金项目
安徽重点研究与开发计划项目(1804a09020009) （1804a09020009）
安徽省高校自然科学研究重大项目与重点项目(J2021ZA0068,KJ2021A1060,2023AH040036) （J2021ZA0068,KJ2021A1060,2023AH040036）
安徽高校协同创新项目(GXXT-2022-019,GXXT2023-006,GXXT-2023-025) （GXXT-2022-019,GXXT2023-006,GXXT-2023-025）
过程装备与控制工程四川省高校重点实验室开放基金项目(GK202101,GK202308).Project Supported by Anhui Provincial Key Research and Development Plan Project,China(1804a09020009),Anhui University Natural Science Research Major Project and Key Project,China(J2021ZA0068,KJ2021A1060,2023AH040036),Anhui University Collaborative Innovation Project,China(GXXT-2022-019,GXXT2023-006,GXXT-2023-025),Process Equipment and Control Engineering Open Fund Project of Key Laboratory of Sichuan University,China(GK202101,GK202308). （GK202101,GK202308）
