基于改进YOLO11的复杂场景带钢表面缺陷检测算法

Complex Scene Strip Steel Surface Defect Detection Algorithm Based on Improved YOLO11

摘要: 针对钢铁表面缺陷检测中存在的形态不规则、特征细微及背景干扰复杂等难题，本文提出一种改进YOLO11的钢铁表面缺陷检测算法。该算法在主干网络的C3k2模块中引入可变形注意力机制，增强模型对不规则缺陷如裂纹(Cr)的空间特征捕捉能力。采用风车形卷积替代部分标准卷积，提升模型对夹杂(Pa)、斑点(Ps)等低对比度细微目标的特征提取性能。同时在C2PSA模块中集成基于变分率削减的线性复杂度注意力机制(TSSA)，有效实现全局上下文信息交互。在NEU-DET数据集上的实验表明，改进模型mAP@0.5达到80.76%，相比原始YOLO11提升3.16%，对比其他的目标检测主流方法，改进算法精度上都有了明显的提升，满足钢铁表面缺陷检测的精度要求。

Abstract: To address challenges in steel surface defect detection, such as irregular shapes, subtle features, and complex background interference, this paper proposes an improved YOLO11 algorithm for steel surface defect detection. The algorithm introduces a deformable attention mechanism in the C3k2 module of the backbone network, enhancing the model’s ability to capture spatial features of irregular defects such as cracks (Cr). Part of the standard convolution is replaced with pinwheel-shaped convolutions to improve the model’s feature extraction performance for low-contrast, fine targets such as inclusions (Pa) and spots (Ps). Additionally, the C2PSA module integrates a linear-complexity attention mechanism (TSSA) based on variance reduction, effectively enabling global context information interaction. Experiments on the NEU-DET dataset show that the improved model achieves an mAP@0.5 of 80.76%, an increase of 3.16% compared to the original YOLO11. Compared with other mainstream object detection methods, the improved algorithm demonstrates a significant improvement in accuracy, meeting the precision requirements for steel surface defect detection.