摘要
针对当前带钢表面缺陷检测领域存在的模型参数量大、检测精度不足以及在工厂实际应用中检测终端性能受限等问题,提出了一种新型轻量级的带钢表面缺陷检测算法YOLOv8n-VAS。首先,在YOLOv8n主干网络中引入VanillaNet模块,该模块通过简化传统分支结构并去除冗余路径,有效减小网络的复杂度,从而实现整体模型的轻量化;其次,在颈部网络中引入AKConv模块,利用不规则卷积操作提升特征提取的自适应能力,显著增强了模型对缺陷关键特征的捕获效果;最后,在C2f模块中引入SCSA注意力机制,通过结合通道与空间注意力在保持轻量化的同时增强特征提取能力。在NEU-DET数据集上的实验结果表明,YOLOv8n-VAS的检测准确率达到了76.4%,参数量和计算量分别为1.87M和5.5G,相比原始YOLOv8n分别降低了32.1%和37.9%,在精度与模型轻量化方面实现了较大幅度地提升。同时,在GC10-DET数据集上进行的泛化实验中,缺陷检测准确率为70.7%,表明YOLOv8n-VAS在不同数据集上均具有良好的检测性能和较强的泛化能力。该方法为工业生产中的质量检测提供了一种高效、轻量化的解决方案。
Aiming at the problems of large model parameters,insufficient detection accuracy,and limited performance of detection terminals in actual factory applications in the field of steel strip surface defect detection,a new lightweight steel strip surface defect detection algorithm YOLOv8n-VAS is proposed.Firstly,the VanillaNet module is introduced into the YOLOv8n backbone network.By simplifying the traditional branch structure and removing redundant paths,the complexity of the network is effectively reduced,thus realizing the lightweighting of the overall model.Secondly,the AKConv module is introduced into the neck network.The irregular convolution operation enhances the adaptive ability of feature extraction and significantly improves the model's ability to capture key defect features.Finally,the SCSA attention mechanism is introduced into the C2f module.By combining channel and spatial attention,the feature extraction ability is enhanced while maintaining lightweighting.The experimental results on the NEU-DET dataset show that the detection accuracy of YOLOv8n-VAS reaches 76.4%,with parameters and calculations of 1.87M and 5.5G respectively,which are reduced by 32.1%and 37.9%compared with the original YOLOv8n.Significant improvements have been made in both accuracy and model lightweighting.At the same time,in the generalization experiment on the GC10-DET dataset,the defect detection accuracy is 70.7%,indicating that YOLOv8n-VAS has good detection performance and strong generalization ability on different datasets,providing an efficient and lightweight solution for quality inspection in industrial production.
作者
刘贺
黄庆尧
安超
钱俊磊
Liu He;Huang Qingyao;An Chao;Qian Junlei(Tangshan Huitang IOT Technology Co.,Ltd.,Tangshan 063009,Hebei;College of Electrical Engineering,North China University of Science and Technology,Tangshan 063210,Hebei)
出处
《河北冶金》
2026年第1期72-80,共9页
Hebei Metallurgy
基金
唐山市科技计划项目(22130204G)。
