基于改进YOLOv12的金属表面缺陷检测算法

Metal Surface Defect Detection Algorithm Based on Improved YOLOv12

摘要: 针对金属表面缺陷检测中存在的缺陷边缘模糊、小目标漏检以及模型收敛效果差等问题，提出了一种基于改进YOLOv12的金属表面缺陷检测算法SCA-YOLO。设计边缘增强卷积模块(EEConv)，通过多方向差分卷积提取缺陷边缘特征，增强模型对划痕、裂纹等线性缺陷的特征表达能力；引入空间通道协同注意力机制(SCA)，通过空间注意力定位缺陷区域并利用通道注意力筛选重要特征，提升小目标检测精度；引入Wise-IoU损失函数替代传统的CIoU，利用其动态非单调聚焦机制平衡不同质量样本的权重，加速模型收敛。在NEU-DET数据集上的实验结果表明，SCA-YOLO的mAP@0.5达到78.9%，相比基线YOLOv12n提升5.7%，同时参数量仅增加3.5%，验证了所提方法的有效性。

Abstract: Aiming at the problems of fuzzy defect edges, missed detection of small targets, and poor model convergence in metal surface defect detection, an improved YOLOv12-based algorithm named SCA-YOLO is proposed. An Edge Enhancement Convolution (EEConv) module is designed to extract defect edge features through multi-directional differential convolution, enhancing the model’s feature representation capability for linear defects such as scratches and cracks. A Spatial-Channel Collaborative Attention (SCA) mechanism is introduced, which locates defect regions via spatial attention and filters important features through channel attention, thereby improving the detection accuracy for small targets. The Wise-IoU loss function replaces the traditional CIoU, utilizing its dynamic non-monotonic focusing mechanism to balance the weights of samples of different qualities and accelerate model convergence. Experimental results on the NEU-DET dataset show that SCA-YOLO achieves an mAP@0.5 of 78.9%, which is 5.7% higher than the baseline YOLOv12n, while the number of parameters increases by only 3.5%, verifying the effectiveness of the proposed method.