用于超纤革表面瑕疵识别的MFL_YOLOv8算法  

MFL_YOLOv8 algorithm for surface defect detection of microfiber leather

作　　者：孙小栋 朱启兵[1] 徐华伟 邢同振 朱海斌 SUN Xiaodong;ZHU Qibing;XU Huawei;XING Tongzhen;ZHU Haibin(School of Internet of Things Engineering,Jiangnan University,Wuxi 214122,China;Hexin Kuraray Micro Fiber Leather(Jiaxing)Co.,Ltd,Jiaxing 314003,China;Zhejiang Maimu Intelligent Technology Co.,Ltd,Jiaxing 314000,China)

机构地区：[1]江南大学物联网工程学院,江苏无锡214122 [2]禾欣可乐丽超纤皮(嘉兴)有限公司,浙江嘉兴314003 [3]浙江迈沐智能科技有限公司,浙江嘉兴314000

出　　处：《光学精密工程》2025年第2期311-323,共13页Optics and Precision Engineering

基　　金：浙江省基础公益研究计划项目(No.LZ22A020006)。

摘　　要：超纤革是一种用于高端产品的新型复合材料,其瑕疵检测对产品质量至关重要。针对超纤革表面瑕疵多尺度、长宽比差异大和微小瑕疵较多的难点,提出用于超纤革表面瑕疵识别的MFL_YOLOv8算法。MFL_YOLOv8算法首先基于Deformable Large Kernel Attention(DLKA)机制设计了多尺度特征提取模块DCNv3-LKA,显著增强了主干网络的多尺度特征提取能力;然后通过在特征金字塔网络中引入P2特征图和Dysample上采样模块,强化了网络对小目标的细节信息提取;最后引入Minimum Points Distance Intersection over Union(MPDIoU)以缓解训练初期小目标上的损失函数失效问题,提升了小目标的检测效果。在自制超纤革表面瑕疵数据集上的实验结果表明,相比于YOLOv8n,所提算法的平均检测精度和召回率分别提高了5.38%和7.27%,达到92.47%和92.40%,每秒帧率(FPS)为135.2 frame/s,满足工业现场的准确性和实时性要求。Microfiber leather is a high-end composite material,and its defect detection is critical for ensuring product quality.To address the challenges posed by the multi-scale,diverse aspect ratios,and numerous small defects on the surface of microfiber leather,the MFL_YOLOv8 algorithm for surface defect detection was proposed in this study.The MFL_YOLOv8 algorithm first introduced the multi-scale feature extraction module DCNv3-LKA based on the Deformable Large Kernel Attention(DLKA)mechanism,which significantly enhanced the backbone network's multi-scale feature extraction capabilities.Subsequently,the incorporation of a P2 feature map and a Dysample upsampling module in the feature pyramid network strengthened the network's ability to extract detail information from small targets.Finally,the Minimum Points Distance Intersection over Union(MPDIoU)was utilized to mitigate the inefficacy of the loss function on small targets during the initial stages of training,thus improving the detection performance for small targets.Experimental results on a self-constructed microfiber leather surface defect dataset demonstrate that the proposed algorithm achieved 92.47%of average detection precision and 92.40%of average detection recall,with improvements of 5.38%and 7.27%compared to YOLOv8n.Additionally,the algorithm attainsed a frame rate of 135.2 frames per second(FPS),meeting the accuracy and real-time requirements for industrial applications.

关 键 词：超纤革 瑕疵检测 DCNv3-LKA MPDIoU 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]