基于改进YOLOv8s的钢材表面缺陷检测算法  

Steel surface defect detection algorithm based on improved YOLOv8s

作　　者：梁礼明 龙鹏威 金家新 李仁杰 曾璐 LIANG Liming;LONG Pengwei;JIN Jiaxin;LI Renjie;ZENG Lu(School of Electrical Engineering and Automation,Jiangxi University of Science and Technology,Ganzhou 341000,China)

机构地区：[1]江西理工大学电气工程与自动化学院,江西赣州341000

出　　处：《浙江大学学报(工学版)》2025年第3期512-522,共11页Journal of Zhejiang University:Engineering Science

基　　金：国家自然科学基金资助项目(51365017,61463018);江西省自然科学基金资助项目(20192BAB205084);江西省教育厅科学技术研究资助项目(GJJ2200848).

摘　　要：钢材表面缺陷形态多样、结构复杂、小目标占比高,而通用目标检测算法计算量过大且不适合终端设备部署.针对上述问题,提出基于YOLOv8s的轻量级的钢材缺陷检测算法(SDB-YOLOv8s).重新设计特征交互模块(S-C2f),抑制空间和通道冗余信息,提高检测精度;引入空洞Transformer模块,增强网络对全局上下文信息的捕获能力和稀疏采样特性,以减少细粒度信息损失,并提升对小目标的特征提取能力;设计BS-ShuffleNetV2轻量化网络作为骨干网络,在降低模型复杂度的同时保证检测精度.在NEU-DET和Severstal钢材缺陷数据集进行实验验证,结果表明,与基线模型相比,SDB-YOLOv8算法的mAP分别提升6.4和7.0个百分点、参数量和计算复杂度仅为基线模型的64.8%和56.2%.每秒检测帧数分别达到146帧和121帧、精确度分别提升4.6和6.5个百分点.实验结果表明,该算法在检测精度、速度和轻量化方面取得了较好的平衡,同时为边缘终端设备提供了较高精度和实时性的参考.The surface defects of steel are diverse in form,complex in structure,and exhibit a high proportion of small targets,while general object detection algorithms have excessive computational complexity and are not suitable for the deployment on edge devices.To address these issues,a lightweight steel defect detection algorithm based on YOLOv8s,called SDB-YOLOv8s,was proposed.Firstly,a redesigned feature interaction module(S-C2f)was introduced to suppress spatial and channel redundant information,enhancing detection accuracy.Secondly,a dilated Transformer module was incorporated to enhance the network’s ability to capture global contextual information and sparse sampling characteristics,reducing fine-grained information loss and improving feature extraction capabilities for small targets.Finally,a lightweight network,BS-ShuffleNetV2,was designed as the backbone network to reduce model complexity while maintaining detection accuracy.Experimental validation on the NEU-DET and Severstal steel defect datasets showed that compared to baseline models,the SDB-YOLOv8 algorithm achieved an improvement in mAP of 6.4 and 7.0 percentage points,and detection frames per second of 146 and 121,with accuracy improvements of 4.6 and 6.5 percentage points respectively.The number of parameters and computational complexity were only 64.8%and 56.2%of the baseline model.The experimental results demonstrated that this algorithm achieved a better balance in terms of detection accuracy,speed,and lightweight characteristics,while providing a reference for high accuracy and real-time capabilities for edge terminal devices.

关 键 词：缺陷检测 轻量化YOLOv8s 深度学习 特征提取 特征交互 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]