基于改进YOLOv8s的织物疵点检测  

Fabric defect detection based on improved YOLOv8s

作　　者：王帅 刘珊珊 李保田 张永军 陈健健 WANG Shuai;LIU Shanshan;LI Baotian;ZHANG Yongjun;CHEN Jianjian(Shandong Youth University of Political Science,Jinan,250103,China;New Technology Research and Development Center of Intelligent Information Controlling for High Education Institution in Shandong Province,Jinan,250103,China;Smart Healthcare Big Data Engineering and Ubiquitous Computing Characteristic Laboratory for High Education Institution in Shandong Province,Jinan,250103,China;Jinan Vocational College,Jinan,250103,China)

机构地区：[1]山东青年政治学院,山东济南250103 [2]山东省高等学校智能信息控制新技术研发中心,山东济南250103 [3]山东省高等学校智慧康养大数据工程与泛在计算特色实验室,山东济南250103 [4]济南职业学院,山东济南250103

出　　处：《棉纺织技术》2025年第2期49-55,共7页Cotton Textile Technology

基　　金：山东青年政治学院博士科研启动基金(XXPY23036)。

摘　　要：为了解决人工检查效率低、算法检测的缺陷种类少以及算法检测精度差等问题,提出了一种改进的YOLOv8s算法来进行织物疵点检测。首先,通过将DCNv4算子融入YOLOv8s特征融合网络Neck中的C2f模块,构建了C2f_DCNv4模块,不仅提高了对小目标的检测精度,还实现了比使用DCNv3更快的处理速度。其次,将MSCA注意力模块与常见的CBAM注意力相结合,创新设计了MCASAM注意力模块,并将其加入到特征提取网络中的最后一层,提升了模型的多尺度特征捕获能力以及对小目标的检测敏感度。最后,在损失函数中,用InnerIoU替换了传统的CIoU,不仅收敛速度更快,而且进一步优化了检测框的精确度。试验结果表明:在天池疵点检测数据集上针对20类疵点的对比试验中,与原始YOLOv8s算法相比,该研究所提方法的mAP@0.5值达到0.608,提高了5.9%,模型推理速度达150帧/s,能够满足实际的检测需求。In order to solve the problems of lower manual inspection efficiency,few types of defects in the dataset,and poor algorithm detection accuracy,an improved YOLOv8s algorithm for fabric defect detection was proposed.Firstly,by integrating DCNv4 operator into the C2f module of the YOLOv8s feature fusion network Neck,C2f_DCNv4 module was constructed,which could not only improve the detection accuracy of small targets but also achieve faster processing speed than DCNv3.Secondly,by combining the MSCA attention module with common CBAM attention,the MCASAM attention module was innovatively designed and added to the last layer of the feature extraction network,improving the model's multi-scale feature capture ability and sensitivity to detecting small targets.Finally,in the loss function,the traditional CIoU was replaced by InnerIoU,which not only converged faster but also further optimized the accuracy of the detection box.The test results showed that in the comparative experiment of 20 types defects on the Tianchi defect detection dataset,compared with the original YOLOv8s,mAP@0.5 value of the proposed method was reached 0.608,which was improved 5.9%,and the model inference speed was reached 150 frames/s,which can meet the practical detection needs.

关 键 词：YOLOv8s DCNv4 MCASAM CBAM InnerIoU 织物疵点检测 C2f_DCNv4 

分 类 号：TS107[轻工技术与工程—纺织工程]