摘要
钢铁行业是我国的支柱产业,钢材的质量是影响钢材性能与价格的关键。针对带钢表面缺陷检测存在精度差、效率低以及模型结构复杂等问题,提出一种基于YOLOv11的轻量化带钢表面缺陷检测算法(PSN-YOLO)。首先设计P-GELAN_CAA特征提取-融合模块,基于广义高效层聚合网络(GELAN)引入多尺度卷积(PSConv)处理多尺度信息、优化参数利用率,并融合上下文锚点注意力(CAA)增强特征表示;其次,选取轻量高效的SCDown下采样以扩大感受野,减少信息损失,降低模型复杂度;最后,采用归一化瓦瑟斯坦距离(NWD)改进边界框损失函数,专注于不规则和复杂微小纹理特征,更好地衡量边界框间的分布相似性,提升检测精度。在NEU-DET数据集上的试验结果表明,该模型相比基准模型mAP提高了3.1%,参数量和计算量分别减少20.3%、19.0%,更好地平衡了检测精度和轻量化需求。此外,该模型在Severstal数据集上表现出良好的泛化能力,满足实际工程需求,具有重要推广应用价值。
Iron and steel is a pillar industry in China,and the production quality of steel products is key to the performance andprice.In order to solve the problems of poor accuracy,low efficiency and complex model structure in strip surface defect detection,we proposed a lightweight strip surface defect detection algorithm based on YOLOv11(PSN-YOLO).Firstly,the P-GELAN_CAA feature extraction-fusion module was designed,and PSConv was introduced based on GELAN to process multi-scale information,optimize parameter utilization,and integrate CAA to enhance feature representation.Secondly,the lightweight and efficient SCDown downsampling was selected to expand the receptive field,reduce the information loss,and reduce the complexity of the model.Finally,NWD is used to improve the loss function of the bounding box,focusing on irregular and complex micro-texture features,so as to better measure the distribution similarity between the bounding boxes and improve the detection accuracy.Experimental results on the NEU-DET dataset show that compared with the benchmark model,the mAP of the proposed model is increased by 3.1%,and the number of parameters and computation are reduced by 20.3%and 19.0%,respectively,which better balances the detection accuracy and lightweight requirements.In addition,the model shows good generalization ability on the Severstal dataset,which meets the practical engineering needs and has important application value.
作者
马俊杰
张继红
王强
刘文广
吴振奎
MA Junjie;ZHANG Jihong;WANG Qiang;LIU Wenguang;WU Zhenkui(College of Automation and Electrical Engineering,Inner Mongolia University of Science and Technology,Baotou 014010,Nei Mongol,China;Equipment Division,Baotou Wei Feng New Materials Company Limited,Baotou 014020,Nei Mongol,China;College of Mechanical Engineering,Inner Mongolia University of Science and Technology,Baotou 014010,Nei Mongol,China)
出处
《钢铁研究学报》
北大核心
2025年第10期1345-1358,共14页
Journal of Iron and Steel Research
基金
内蒙古自治区重点研发与成果转化资助项目(2022YFHH0019)
内蒙古自治区科技攻关大平台建设资助项目(2023PTXM001)。
