Abstract

Deep learning algorithms have gained widespread usage in defect detection systems. However, existing methods are not satisfied for large-scale applications on surface defect detection of strip steel. In this paper, we propose a precise and efficient detection model, named CABF-YOLO, based on the YOLOX for strip steel surface defects. Firstly, we introduce the Triplet Convolutional Coordinate Attention (TCCA) module in the backbone of the YOLOX. By factorizing the pooling operation, the TCCA module can accurately capture cross-channel features to identify the location information of defects. Secondly, we design a novel Bidirectional Fusion (BF) strategy in the neck of the YOLOX. The BF strategy enhances the fusion of low-level and high-level semantic information to obtain fine-grained information. Lastly, the original bounding box loss function is replaced by the EIoU loss function. In the EIoU loss function, the penalty term is redefined to consider the overlap area, central point, and side length of the required regressions to accelerate the convergence rate and localization accuracy. On the benchmark NEU-DET dataset and GC10-DET dataset, the experimental results show that the CABF-YOLO achieves superior performance compared with other comparison models and satisfies the real-time detection requirement of industrial production.

中文翻译：

---

CABF-YOLO：一种精确高效的带钢表面缺陷检测深度学习方法

摘要

深度学习算法在缺陷检测系统中得到了广泛的应用。然而，现有方法并不能满足带钢表面缺陷检测的大规模应用。在本文中，我们基于YOLOX提出了一种精确高效的带钢表面缺陷检测模型，命名为CABF-YOLO。首先，我们在YOLOX的主干中引入三重卷积坐标注意（TCCA）模块。通过分解池化操作，TCCA模块可以准确捕获跨通道特征来识别缺陷的位置信息。其次，我们在 YOLOX 的颈部设计了一种新颖的双向融合（BF）策略。 BF策略增强了低层和高层语义信息的融合，以获得细粒度的信息。最后，将原来的边界框损失函数替换为EIoU损失函数。在EIoU损失函数中，重新定义了惩罚项，以考虑所需回归的重叠面积、中心点和边长，以加快收敛速度​​和定位精度。在基准NEU-DET数据集和GC10-DET数据集上，实验结果表明CABF-YOLO与其他对比模型相比取得了优越的性能，满足工业生产的实时检测要求。