针对传统钢材表面缺陷检测方法易出现误检、漏检和部分缺陷种类检测精度低等问题，设计了一种钢材表面缺陷网络YOLOv5s-FCS。首先引用了FReLU激活函数构建了卷积模块CBF，有效增强了网络的空间解析能力，优化了网络检测精度；其次，将坐标注意力机制嵌入到网络的neck部分来增强网络特征融合的能力，从而使网络能够提取更加丰富的特征信息；最后，将YOLOv5s的损失函数替换为SIoU loss，提高了预测框的回归精度。通过在NEU-DET数据集上进行消融实验、可视化对比实验，结果表明，YOLOv5s-FCS网络的mAP值达到了0.747，相较于原YOLOv5s网络提高了8.3%，相较于YOLOv3网络提高了11.8%，相较于YOLOXs网络提高了4.2%，相较于YOLOv6s提高了1.4%，验证了该方法的可行性、有效性。

The YOLOv5s-FCS network for traditional steel materials, which addresses issues such as false positives, false negatives, and low accuracy in detecting certain types of defects was presented. Firstly, the CBF convolution module was constructed using the FReLU activation function to enhance the network's spatial resolution capability and optimize detection accuracy. Secondly, a coordinate attention mechanism was embedded into the neck part of the network to enhance its feature fusion capability, enabling the extraction of more rich feature information. Finally, the SIoU loss replaces the YOLOv5s loss function to improve the regression accuracy of the predicted box. Through ablation experiments and visualization comparisons on the NEU-DET dataset, it is demonstrated that the mAP value of the YOLOv5s-FCS network reaches 0.747, representing an improvement of 8.3% compared to the original YOLOv5s network, 11.8% compared to the YOLOv3 network, 4.2% compared to the YOLOXs network, and 1.4% compared to the YOLOv6s network, thus demonstrating the feasibility and effectiveness of the proposed method.