基于RM-YOLOv8的热轧带钢表面缺陷检测  

Surface defect detection on hot rolled strip steel based on RM-YOLOv8

作　　者：杨澳 侯红玲[1] 朱康凯 李相垚 赵妍棣 Yang Ao;Hou Hongling;Zhu Kangkai;Li Xiangyao;Zhao Yandi(School of Mechanical Engineering,Shaanxi University of Technology,Hanzhong 723001,China;School of Mechanical and Precision Instrument Engineering,Xi'an University of Technology,Xi'an 710048,China)

机构地区：[1]陕西理工大学机械工程学院,陕西汉中723001 [2]西安理工大学机械与精密仪器工程学院,陕西西安710048

出　　处：《锻压技术》2025年第2期103-114,共12页Forging & Stamping Technology

基　　金：陕西省自然科学基础研究计划项目(2023-JC-YB-452)。

摘　　要：针对传统热轧带钢表面缺陷检测算法在复杂工业环境下检测精度低、易漏检和误检等问题,提出一种基于RM-YOLOv8的热轧带钢表面缺陷检测算法。在特征提取网络中引入ConvNeXt Block模块,增强了网络特征提取能力;构建了注意力机制RC_MHSA,加强了特征提取网络对于特征图中感兴趣区域的训练;构建了轻量可学习视觉中心模块LLVC,并以跨层通道融合的方式向深层网络中加入更加丰富的特征信息,提高了模型对于局部角落特征的捕捉;最后,利用GhostBottleneck模块构造了一种轻量化模块C2f_Ghost,在减少计算量的同时,提高了网络的检测精度。实验结果表明,RM-YOLOv8在热轧带钢表面缺陷数据集NEU-DET上实验的平均精度均值相比于原始YOLOv8网络提升了7.0%,精确率提高了8.9%,召回率提高了6.8%。与其他检测算法相比,提出的RM-YOLOv8缺陷检测算法具有更好的检测性能和工业应用价值。For the problems of low detection accuracy,easy to miss detection and misdetect in complex industrial environments caused by traditional algorithms for the surface defect detection on hot rolled strip steel,a surface defect detection algorithm of hot rolled strip steel based on RM-YOLOv8 was proposed,and ConvNeXt Block module was introduced into the feature extraction network to enhance the feature extraction capability of the network.Then,attention mechanism RC_MHSA was constructed to enhance the training of feature extraction network for the regions of interest in feature maps,and a lightweight and learnable visual center module(LLVC)was constructed.Furthermore,the richer feature information was added to the deep network by cross-layer channel fusion to improve the model ability to capture local corner features,and finally a lightweight module C2f_Ghost was constructed by using GhostBottleneck module,which reduced the amount of computation and improved the detection accuracy of the network.The experimental results show that the average mean precision of RM-YOLOv8 on the NEU-DET dataset of surface defects in hot rolled strip steel is increased by 7.0%,the accuracy is increased by 8.9%,and the recall rate is increased by 6.8%compared to the original YOLOv8 network.Compared with other detection algorithms,the proposed RM-YOLOv8 defect detection algorithm has better detection performance and industrial application value.

关 键 词：热轧带钢 缺陷检测 深度学习 RM-YOLOv8 特征金字塔 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]