基于FG-YOLOv7-tiny算法的耐张线夹X光图像压接缺陷检测  

Detection of Crimping Defects in Strain Clamp X-ray Images Based on FG-YOLOv7-tiny Algorithm

作　　者：杨宇 高林 唐永欣 王志 廖明艳 YANG Yu;GAO Lin;TANG Yongxin;WANG Zhi;LIAO Mingyan(College of Intelligent Systems Science and Engineering,Hubei Minzu University,Enshi 445000,China;School of Mathematics and Statistics,Chongqing Technology and Business University,Chongqing 400000,China;Enshi Prefecture Productivity Promotion Center,Enshi 445000,China)

机构地区：[1]湖北民族大学智能科学与工程学院,湖北恩施445000 [2]重庆工商大学数学与统计学院,重庆400000 [3]恩施州生产力促进中心,湖北恩施445000

出　　处：《湖北民族大学学报（自然科学版）》2024年第1期51-58,共8页Journal of Hubei Minzu University:Natural Science Edition

基　　金：国家自然科学基金项目(61562025,61962019);湖北省高等学校省级教学研究项目(2017387)。

摘　　要：为保证输电线路的安全可靠运行,电力巡检的重要任务是耐张线夹压接缺陷检测。为此,提出了快速幽灵YOLOv7-tiny(faster neural networks ghost convolution-you only look once version 7-tiny,FG-YOLOv7-tiny)算法进行耐张线夹压接缺陷检测。首先,构建包含5类常见压接缺陷的耐张线夹X光图像数据集;其次,使用快速神经网络(faster neural networks,FasterNet)替代YOLOv7-tiny的高效聚合网络(efficient layer aggregation networks,ELAN)以减小模型大小;最后,使用幽灵空间金字塔池化交叉阶段部分连接网络(ghost spatial pyramid pooling cross stage partial connection networks,GhostSPPCSPC)替换YOLOv7-tiny使用的空间金字塔池化交叉阶段部分连接网络以提升检测精度。实验结果表明,FG-YOLOv7-tiny算法的精度、平均精度均值分别达到91.30%、94.28%,相比于原始YOLOv7-tiny算法分别提升了3.99%、1.59%;模型大小为22.25 MB;检测速度达到172.41帧/s,能满足实时检测的要求。因此,FG-YOLOv7-tiny算法提升了检测精度,可实现耐张线夹压接缺陷的有效检测,并满足边缘设备部署的要求。To ensure the safe and reliable operation of transmission lines,the detection of crimp defects in strain clamps had become an important task of power inspection.To this end,a faster neural networks ghost convolution-you only look once version 7-tiny(FG-YOLOv7-tiny)algorithm was proposed for strain clamp crimp defect detection.First,a strain-clamp X-ray image dataset containing five common crimp defects was constructed.Secondly,faster neural networks(FasterNet)were used to replace the efficient layer aggregation networks(ELAN)of YOLOv7-tiny to reduce the model size.Finally,the ghost spatial pyramid pooling cross stage partial connection networks(GhostSPPCSPC)was used to replace the spatial pyramid pooling cross stage partial connection networks used by YOLOv7-tiny to improve the detection accuracy.The experimental results showed that the accuracy and average accuracy of the FG-YOLOv7-tiny algorithm reached 91.30%and 94.28%,respectively,which were 3.99%and 1.59%higher than the original YOLOv7-tiny algorithm.The model size was 22.25 MB,and the detection speed reached 172.41 frames/s,which satisfied the requirements of real-time detection.Therefore,the FG-YOLOv7-tiny algorithm improved the detection accuracy,which could effectively detect the crimping defects of strain clamps and meet the requirements of edge device deployment.

关 键 词：YOLOv7-tiny 快速神经网络 耐张线夹 缺陷检测 X光图像 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]