Scholarlink

<p>Steel surface defect detection plays a critical role in industrial quality control. However, current methods have difficulty balancing accuracy and speed, especially in complex defect environments, which present significant challenges. To overcome these challenges, this paper proposes the Fast-YOLOv12 surface defect detection algorithm. First, we replace the original backbone with the lightweight FasterNet, enhancing its ability to capture semantic feature information more effectively. Second, we propose the Sage-Area Attention and integrate it into the attention-augmented cross-stage fusion (A2C2f) module, naturally forming SGC2f block. This enhancement boosts both the capability of the model for complex feature extraction and its inference speed. Finally, the Wise-ShapeIoU replaces the traditional CIoU, which improves the precision and stability of target localization. Experimental validation on the NEU-DET dataset shows that Fast-YOLOv12 improves mAP@0.5 by 4.2–79.1% over the baseline model. Fast-YOLOv12 achieves a balance between a lightweight design and real-time performance, while also improving the accuracy of surface defect detection. This enhances its suitability for deployment on edge devices. Moreover, evaluations on the GC10-DET dataset demonstrate the strong generalization ability of the proposed model.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Fast-YOLOv12: an attention-guided lightweight network for real-time steel surface defect detection

  • 期刊论文

  • 发表日期：

    2025年11月17日

查看全文
  • Yanfeng Qiao,
  • Yiru Wei,
  • Kai Wang

摘要

Steel surface defect detection plays a critical role in industrial quality control. However, current methods have difficulty balancing accuracy and speed, especially in complex defect environments, which present significant challenges. To overcome these challenges, this paper proposes the Fast-YOLOv12 surface defect detection algorithm. First, we replace the original backbone with the lightweight FasterNet, enhancing its ability to capture semantic feature information more effectively. Second, we propose the Sage-Area Attention and integrate it into the attention-augmented cross-stage fusion (A2C2f) module, naturally forming SGC2f block. This enhancement boosts both the capability of the model for complex feature extraction and its inference speed. Finally, the Wise-ShapeIoU replaces the traditional CIoU, which improves the precision and stability of target localization. Experimental validation on the NEU-DET dataset shows that Fast-YOLOv12 improves mAP@0.5 by 4.2–79.1% over the baseline model. Fast-YOLOv12 achieves a balance between a lightweight design and real-time performance, while also improving the accuracy of surface defect detection. This enhances its suitability for deployment on edge devices. Moreover, evaluations on the GC10-DET dataset demonstrate the strong generalization ability of the proposed model.