基于YOLOv8s的特征融合织物疵点检测方法  

Fabric defect detection method based on improved YOLOv8s

作　　者：蒋明宇 吴斌[1] 张红英[1] JIANG Mingyu;WU Bin;ZHANG Hongying(School of Information Engineering,Southwest University of Science and Technology,Mianyang,Sichuan 621000,China)

机构地区：[1]西南科技大学信息工程学院,四川绵阳621000

出　　处：《毛纺科技》2025年第3期137-144,共8页Wool Textile Journal

摘　　要：针对织物缺陷种类繁多、尺度变化大以及小目标漏检等问题,提出了一种基于YOLOv8s的特征融合改进算法。首先在Neck层引入加权特征融合策略,并设计了多分辨率特征层聚合模块CWcat,通过加权融合不同尺度的特征信息,显著提升了检测精度,同时仅增加少量计算成本;其次,采用VOVDGSCSP特征融合模块优化了Neck层的多层特征信息融合,提高了小目标检测的精度,并有效降低了计算复杂度;最后,采用SIoU损失函数替代原有的边框损失函数,增强了类别分类能力和训练收敛性,提高了模型的稳定性。实验结果表明:改进后的算法在天池织物数据集上取得了86.0%的mAP@0.5、88.5%的精确率和78.5%的召回率,分别较原YOLOv8s提升3.3%、2.1%和3.0%;参数量和计算量分别降低了5.5%和3.4%。该改进算法在织物缺陷检测中表现出显著精度和效率提升,能更好地应对实际生产中的检测挑战。An improved feature fusion algorithm based on YOLOv8s was proposed to solve the problems of fabric defects,large scale variation and small target missing detection.Firstly,weighted feature fusion strategy was introduced into the Neck layer,and multi-resolution feature layer aggregation module CWcat was designed.By weighted fusion of feature information of different scales,the detection accuracy was significantly improved,and the calculation cost was only increased a little.Secondly,VOVDGSCSP feature fusion module was used to optimize the multi-layer feature information fusion of Neck layer,which improves the precision of small target detection and effectively reduces the computational complexity.Finally,SIoU loss function was used to replace the original frame loss function,which enhances the class classification ability and training convergence,and improves the stability of the model.Experimental results show that the improved algorithm achieves 86.0% mAP@0.5,88.5% accuracy and 78.5% recall rate on Tianchi fabric dataset,which are 3.3%,2.1% and 3.0% higher than the original YOLOv8s,respectively.The number of parameters and the amount of computation decreased by 5.5% and 3.4%,respectively.The improved algorithm shows significant improvement in precision and efficiency in fabric defect detection,and can better meet the detection challenges in actual production.

关 键 词：织物疵点 YOLOv8s 加权特征融合 GSConv VOVGSCSP SIoU 

分 类 号：TS107[轻工技术与工程—纺织工程]