基于STCS-YOLO的带钢表面缺陷检测算法  

Defect detection algorithm of strip surface based on STCS-YOLO

作　　者：周亚罗 武献超 刘文广[2] 张瑞成 ZHOU Yaluo;WU Xianchao;LIU Wenguang;ZHANG Ruicheng(Collage of Electrical Engineering,North China University of Science and Technology,Tangshan 063210,Hebei,China;Shougang Jingtang Limited Iron and Steel Co.,Ltd.,Tangshan 063200,Hebei,China)

机构地区：[1]华北理工大学电气工程学院,河北唐山063210 [2]首钢京唐钢铁联合有限责任公司,河北唐山063200

出　　处：《中国冶金》2023年第12期128-138,共11页China Metallurgy

基　　金：河北省自然科学基金资助项目(F2018209201);唐山市科技局科技计划项目(22130213G)。

摘　　要：带钢表面缺陷检测技术是高质量带钢产品生产的重要技术之一。针对以往带钢表面缺陷检测中存在漏检、定位不准、小尺度缺陷目标检测能力较差的问题,提出了一种改进YOLOv5的带钢表面缺陷检测算法(STCS-YOLO)。首先,在特征融合网络的输出部分采用Swin Transformer模块与原有C3模块相融合,增强对全局特征信息的交互与复用,显著提高了对小尺度缺陷目标的检测能力;其次,采用一种轻量级上采样算子CARAFE来替换传统上采样操作,以更好地恢复缺陷信息,提高对带钢表面缺陷的识别精度;最后,在特征提取网络中嵌入3-D权值注意力机制SimAM,以加强对前景特征信息的关注能力,提高对缺陷目标的强辨识能力。试验结果表明,所提算法在NEU-DET数据集上均值平均精度PmA达到了79.7%,比原网络提高了3.9个百分点,并且在模型权重与计算复杂度几乎不变的情况下,单帧检测时间达到了10.9 ms,基本能够满足带钢表面缺陷准确、快速的检测需求。本研究提出的带钢表面缺陷检测算法为生产整洁、无瑕的高质量带钢产品奠定了技术基础。Strip steel surface defect detection technology is one of the important technologies for the production of high quality strip steel products.Aiming at the problems of misdetection,inaccurate positioning,and poor detection ability of small-scale defect targets in previous strip surface defect detection,a strip surface defect detection algorithm(STCS-YOLO)of improved YOLOv5 was proposed.Firstly,the Swin Transformer module was used to fuse with the original C3 module in the output part of feature fusion network to enhance the interaction and reuse of global feature information,which significantly improved the detection ability of small-scale defect targets.Secondly,a lightweight upsampling operator CARAFE was used to replace the traditional sampling operation to better recover defect information,which improved the recognition accuracy of strip surface defects.Finally,embedding 3-D weight attention mechanism SimAM in feature extraction network to enhance the ability of focusing on the foreground feature information,which improved the strong identification ability of defect target.The experimental results demonstrate that the proposed algorithm achieves 79.7%of PmA(mean average precision)on the NEU-DET dataset,which is improvement of 3.9 percentage points over the original network.Additionally,while maintaining model weight and computational complexity nearly unchanged,the single-frame detection time reaches 10.9 ms,which can basically meet the requirements of accurate and rapid detection of strip surface defects.The strip surface defect detection algorithm proposed in this paper lays the technological foundation for improving neat and flawless high quality strip products.

关 键 词：表面缺陷检测 YOLOv5 Swin Transformer CARAFE 注意力机制 

分 类 号：TG115[金属学及工艺—物理冶金] TP183[金属学及工艺—金属学] TP391.41[自动化与计算机技术—控制理论与控制工程]