Abstract: An improved YOLOv8n detection model is proposed to address the issue of insufficient accuracy in steel surface defect detection. Firstly, the C2f module in the Head network is redesigned using deformable convolutions, effectively enhancing the detection capability for irregular surface defects by leveraging the variability of sampling points. Secondly, a multi-head self-attention mechanism is introduced to capture diverse focal points, providing comprehensive and diversified feature representations, further enhancing the  detection performance of model. Finally, the features of the Backbone and Head are concatenated to enhance the quality and richness of the features, solving the problem of local information loss of defects. Experimental validation on the NEU-DET steel strip dataset demonstrates that compared to the original YOLOv8n model, this approach achieves an increase of 5.6?percentage points in average precision and 2.2?percentage points in mAP50. Notably, significant improvements of 15.2?percentage points in precision and 9.9 percentage points in mAP50 are observed in the detection of crack defects.

Key words: object detection, surface defects on steel strips, deformable convolution, YOLOv8

摘要： 针对带钢表面缺陷检测精度不足的问题，提出一种改进的YOLOv8n检测模型。基于可变形卷积对Head网络中的C2f模块进行了重新设计；通过利用采样点的可变性，有效提高了对表面不规则缺陷的检测能力。添加多头自注意力机制捕捉不同的关注点，从而提供全面且多样化的特征表示，进一步增强模型的检测性能。拼接Backbone和Head的特征，提升特征的质量和丰富度，解决缺陷局部信息丢失问题。在NEU-DET带钢数据集的实验验证中，该方法在平均精度和mAP50上较YOLOv8n分别提升了5.6和2.2个百分点。值得注意的是，裂纹缺陷的检测精度和mAP50分别提升了15.2和9.9个百分点，效果显著。

关键词: 目标检测, 表面缺陷, 可变形卷积, YOLOv8