Scholarlink

<p>To address the issues of false positives and missed detections caused by complex-background interference, difficulties in target feature extraction, and insufficient recognition accuracy for irregular defects in steel surface defect detection, an efficient detection algorithm named SFC-DETR based on an improved RT-DETR framework is proposed. First, a lightweight backbone network, StarNet, is introduced to reduce computational complexity while enhancing multi-scale feature extraction capability. Second, a Fusion-GELAN module is designed by integrating nested cross-stage connections with a channel attention mechanism to improve feature representation. Then, an improved CGBlock module is incorporated to achieve progressive downsampling, effectively preserving feature information of various defect targets. Finally, the combination of SIoU angle-aware loss and a Slide-VFL hard-sample weighting strategy is adopted to further optimize detection performance. Experimental results on the NEU-DET dataset demonstrate that, while maintaining detection speed, the proposed model reduces the number of parameters by 32% and computational cost by 44%, achieving a mean average precision (mAP) of 76.1%, which is an improvement of 3.6% over the baseline model. The proposed method achieves accurate, efficient, and compact steel defect detection for reliable industrial inspection.</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

SFC‑DETR: a lightweight transformer with adaptive multi-scale feature fusion for real-time steel surface defect detection

  • 期刊论文

  • 发表日期：

    2026年6月26日

查看全文
  • Jiuhao Ruan,
  • Zhaolin Cheng,
  • Mengjia Zeng

摘要

To address the issues of false positives and missed detections caused by complex-background interference, difficulties in target feature extraction, and insufficient recognition accuracy for irregular defects in steel surface defect detection, an efficient detection algorithm named SFC-DETR based on an improved RT-DETR framework is proposed. First, a lightweight backbone network, StarNet, is introduced to reduce computational complexity while enhancing multi-scale feature extraction capability. Second, a Fusion-GELAN module is designed by integrating nested cross-stage connections with a channel attention mechanism to improve feature representation. Then, an improved CGBlock module is incorporated to achieve progressive downsampling, effectively preserving feature information of various defect targets. Finally, the combination of SIoU angle-aware loss and a Slide-VFL hard-sample weighting strategy is adopted to further optimize detection performance. Experimental results on the NEU-DET dataset demonstrate that, while maintaining detection speed, the proposed model reduces the number of parameters by 32% and computational cost by 44%, achieving a mean average precision (mAP) of 76.1%, which is an improvement of 3.6% over the baseline model. The proposed method achieves accurate, efficient, and compact steel defect detection for reliable industrial inspection.