基于YOLOv5改进模型的金属表面缺陷检测实验研究  

Experimental study on metal surface defect detection using an improved YOLOv5 model

作　　者：赵辉 陈志峰[1] 章佳伟 李骁凡 魏震杨 ZHAO Hui;CHEN Zhifeng;ZHANG Jiawei;LI Xiaofan;WEI Zhenyang(School of Physics and Materials Science,Guangzhou University,Guangzhou 510006,China;Guangdong Tianshi Intelligent Technology Co.,Ltd.,Foshan 528041,China)

机构地区：[1]广州大学物理与材料科学学院,广东广州510006 [2]广东天视智能科技有限公司,广东佛山528041

出　　处：《实验技术与管理》2025年第1期66-74,共9页Experimental Technology and Management

基　　金：广东省高等教育教学改革项目(2019-N466);广州市高等教育教学质量与教学改革工程-实践教学示范中心项目(2023SJJXZX001);广州大学2024年大学生创新训练项目(XJ202411078184)。

摘　　要：金属零部件的表面缺陷检测是汽车等产品生产过程中的重要环节,以往采用人工检视或传统光学筛选方法,该方法难以满足现代工业生产的高效性和准确性要求。该研究源于校企协同育人的创新课题,从企业提出的实际问题出发,选取汽车喷油管表面缺陷作为研究案例。为了提高对弱小缺陷的检测准确率和速度,提出了一种基于YOLOv5轻量化模型的改进结构:YOLOv5n-STSL。该模型通过改进原模型中的卷积模块C3为C2f模块,在保证轻量化的同时获取了更丰富的梯度信息流;通过往浅层特征图移动检测分支,增加不同层次的特征融合,提高了弱小目标的特征提取和检测能力;同时改进锚框anchors的计算评估策略,确保锚框与真实缺陷的边界框有更高的匹配精度,从而提高定位和分类的准确性。实验表明,缺陷检测精度达到97.8%,相对于原基础模型YOLOv5n,检测精度提高了5%,最后将模型部署到嵌入式设备JestonNano,采用TensorRT推理引擎加速推理实验,帧速可达21帧/s,更好地满足了金属表面缺陷自动检测实时性的应用需求。[Objective]This study aims to develop a more efficient and accurate method for detecting surface defects in metal components,which is crucial in industrial production,especially in fields such as automotive manufacturing.Traditional defect detection methods,which often rely on manual visual inspection or optical screening techniques,are inefficient,error-prone,and inadequate for meeting the high standards of modern industry.These methods are particularly ineffective in identifying small or subtle defects,which can significantly impact product quality and reliability.To overcome these limitations,this research focuses on enhancing the YOLOv5 model to improve surface defect detection in automobile fuel injection pipes,a component where precise defect detection is essential.[Methods]This research proposes an enhanced version of the YOLOv5 model,called YOLOv5n-STSL,which integrates several key improvements.The first modification replaces the original C3 convolutional module with the C2f module.This change maintains the model’s lightweight structure while improving its ability to capture richer gradient information,which is essential for detecting small,subtle defects.Additionally,the model detection branch is shifted to shallower feature layers,enabling more effective feature extraction and fusion across different layers,particularly for small and weak targets.Another key improvement lies in the anchor frame calculation strategy,where the traditional evaluation method is replaced with a more accurate approach to ensure better alignment between the anchor boxes and the true boundaries of the defects.This enhanced strategy significantly improves the model’s ability to localize and classify defects more precisely.The model is trained on a large dataset of defect images,and its performance is evaluated based on accuracy,precision,and speed.Finally,the model is deployed on an embedded system,the NVIDIA Jetson Nano,with the TensorRT inference engine used to optimize real-time performance.[Results]Experimental results s

关 键 词：缺陷检测 模型改进 YOLOv5n-STSL ANCHORS 设备部署 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]