电子测量与仪器学报

张新君,赵志闯.基于UAL-YOLO的钢材表面缺陷检测算法[J].电子测量与仪器学报,2026,40(6):126-137

基于UAL-YOLO的钢材表面缺陷检测算法

Research on surface defect detection of steel strip based on UAL-YOLO

DOI：

英文关键词:YOLOv11n defect detection feature fusion lightweighting

基金项目:辽宁省教育厅基本科研项目面上项目（LJKMZ20220678）、辽宁省教育厅重点项目（LJ212410147003）资助

作者	单位
张新君	1.辽宁工程技术大学电子与信息工程学院葫芦岛125105； 2.辽宁省无线射频大数据智能应用重点实验室葫芦岛125105
赵志闯	辽宁工程技术大学电子与信息工程学院葫芦岛125105

Author	Institution
Zhang Xinjun	1.School of Electronic and Information Engineering, Liaoning Technical University, Huludao 125105, China; 2.Liaoning Key Laboratory of Radio Frequency and Big Data for Intelligent Applications, Huludao 125105, China
Zhao Zhichuang	School of Electronic and Information Engineering, Liaoning Technical University, Huludao 125105, China

摘要点击次数: 70

全文下载次数: 15

中文摘要:

针对现有钢材表面缺陷检测模型参数量大，推理速度慢，且模型轻量化与检测性能难以兼顾的问题，提出一种基于YOLOv11n的UAL-YOLO模型。首先，在主干网络末端引入非参数化的统一三维注意力模块（unified 3D attention module,UAM），该模块通过解析特征分布的二阶统计特性动态推导出三维权重，利用神经元间的线性可分性精准捕捉特征空间中的缺陷信息，显著提升了关键语义特征的信噪比。其次，设计了全新的注意力引导融合（attention-guided fusion,AGF）颈部网络，利用深层特征生成引导权重，通过智能和选择性地聚合多尺度特征，增强了模型的特征融合效率与精度。最后，设计轻量级高效偏移上采样器（lightweight efficient offset-based upsampler,LEO）用来替代传统上采样算子，通过LEO模块来学习内容感知的采样点偏移，有效捕捉如细长裂纹边缘、微小麻点坑洼等微弱的纹理特征，提升了模型对低对比度细碎边缘的特征捕捉精度，同时降低了微小缺陷在复杂背景下的漏检率。实验表明，与基础的YOLOv11模型相比较，在NEU-DET数据集上mAP@0.5提升了2.1%，召回率提升了3.7%，同时模型的参数量、计算量和占用储存容量分别下降了26.36%、15.87%和23.64%。UAL-YOLO模型在提升精度的同时实现了显著的轻量化，为高精度、低开销的钢材缺陷检测技术在工业场景的部署提供了新的解决方案。

英文摘要:

Addressing the challenges of large parameter sizes, slow inference speeds, and the difficulty in balancing model lightweighting with detection performance in existing steel surface defect detection models, this paper proposes a UAL-YOLO model based on YOLOv11n. Firstly, a parameter-free unified 3D attention module (UAM) is integrated into the backbone. By deriving 3D weights from second-order statistics, it identifies outlier defect features via linear separability, significantly enhancing the signal-to-noise ratio of semantic information. Secondly, a novel attention-guided fusion (AGF) neck network is designed, which utilizes deep features to generate guiding weights, intelligently and selectively aggregating multi-scale features to improve the efficiency and accuracy of feature fusion. Finally, a lightweight efficient offset-based up sampler (LEO) is developed using content-aware offsets to recover fine textures of slender cracks and minute pits. This approach improves the extraction of low-contrast edges and effectively minimizes the missed detection rate of tiny defects in industrial scenarios. Experimental results demonstrate that compared to the baseline YOLOv11 model, the UAL-YOLO model achieves a 2.1% improvement in mAP@0.5 and a 3.7% increase in recall rate on the NEU-DET dataset, while reducing the number of parameters, computational load, and storage capacity by 26.36%, 15.87%, and 23.64%, respectively. The UAL-YOLO model significantly enhances precision while achieving notable lightweighting, providing a new solution for the deployment of high-precision, low-cost steel defect detection technologies in industrial scenarios.

查看全文查看/发表评论下载PDF阅读器