基于正则化YOLO的钢表面缺陷检测方法

钢铁表面常常显示出错综复杂的纹理模式,这些模式与缺陷相似,给准确识别实际缺陷带来挑战。该研究在基线模型YOLOv8s的基础上提出一种基于正则化YOLO框架的钢表面缺陷检测方法。首先,在C2F框架中嵌入了坐标注意力(CA),利用轻量级注意力模块增强骨干网络的特征提取能力。其次,颈部设计采用可变形卷积(DCN)来加权融合多尺度特征图,增强特征融合能力。最后,对模型的损失函数进行正则化,提高模型的泛化性能。模型在NEU-DET数据集上达到77.94%的mAP0.5。相比基线模型提升2.39%。事实证明该方法更适用于工业检测。

Steel surfaces often display intricate texture patterns that are similar to defects,posing a challenge to accurately identify actual defects.In this study,a steel surface defect detection method based on the regularised YOLO framework is proposed based on the baseline model YOLOv8s.Firstly,coordinate attention(CA)is embedded in the C2F framework to enhance the feature extraction capability of the backbone network using a lightweight attention module.Secondly,the neck design employs deformable convolution(DCN)to weight the fusion of multi-scale feature maps to enhance the feature fusion capability.Finally,the loss function of the model is regularised to improve the generalisation performance of the model.The model achieves 77.94%mAP0.5 on the NEU-DET dataset.a 2.39%improvement over the baseline model.The method proved to be more suitable for industrial inspection.