Scholarlink

<p>In modern industrial production, surface defect detection is essential. Traditional manual methods are inefficient and prone to high false positives. While deep learning has been applied, small sample sizes remain a challenge. To address this, this paper proposes an edge-cloud continuum surface defect detection model based on transfer learning. The architecture spans the cloud, edge, and device layers. By using pre-trained model weights from a cloud repository and fine-tuning with a small dataset at the edge layer, the model reduces training time and reliance on large annotated datasets. Additionally, the YOLOv10-DSC algorithm is introduced, replacing the C2f module with the SCConv module and the original detection head with the Dynamic Head, improving feature extraction and detection efficiency. Experimental results on hot-rolled strip steel, rail, and aluminum sheet defect datasets show that transfer learning accelerates convergence and improves accuracy. The YOLOv10-DSC algorithm outperforms YOLOv10 in mAP, recall, and precision. Specifically, using YOLOv10-DSC with transfer learning from hot-rolled strip defect detection improved mAP@50, mAP@50-95, recall, and precision by 2.8%, 3.9%, 3.6%, and 6%, respectively, compared to direct training with YOLOv10. These results demonstrate the effectiveness of the proposed method, offering a reliable solution for industrial defect detection.</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

An improved transfer learning for surface defect detection in edge-cloud continuum environments

  • 期刊论文

  • 发表日期：

    2026年1月14日

查看全文
  • Bing Tang,
  • Yanming Lin,
  • Yujun Cao,
  • Li Zhang,
  • Xiaoming Ma,
  • Qing Yang

摘要

In modern industrial production, surface defect detection is essential. Traditional manual methods are inefficient and prone to high false positives. While deep learning has been applied, small sample sizes remain a challenge. To address this, this paper proposes an edge-cloud continuum surface defect detection model based on transfer learning. The architecture spans the cloud, edge, and device layers. By using pre-trained model weights from a cloud repository and fine-tuning with a small dataset at the edge layer, the model reduces training time and reliance on large annotated datasets. Additionally, the YOLOv10-DSC algorithm is introduced, replacing the C2f module with the SCConv module and the original detection head with the Dynamic Head, improving feature extraction and detection efficiency. Experimental results on hot-rolled strip steel, rail, and aluminum sheet defect datasets show that transfer learning accelerates convergence and improves accuracy. The YOLOv10-DSC algorithm outperforms YOLOv10 in mAP, recall, and precision. Specifically, using YOLOv10-DSC with transfer learning from hot-rolled strip defect detection improved mAP@50, mAP@50-95, recall, and precision by 2.8%, 3.9%, 3.6%, and 6%, respectively, compared to direct training with YOLOv10. These results demonstrate the effectiveness of the proposed method, offering a reliable solution for industrial defect detection.