Scholarlink

<p>The shape, location, and size of surface defects on steel determine the quality of alloy products. Clearly, implementing automatic segmentation methods for surface defects is valuable. However, given the complexity of the surface defect segmentation task, achieving both segmentation accuracy and real-time prediction in intelligent methods is challenging, which limits their application prospects in the industry. To address the issue, we propose a lightweight deep learning model for defect segmentation tasks. It is a multi-scale feature fusion network (MFF-Metal) that captures effective feature representations through a feature fusion module and a multi-scale attention decoder. Moreover, a fusion loss function and soft label knowledge distillation are designed to improve performances while achieving lightweight prediction. In multiple experiments, compared with multiple existing deep learning models, our method achieves a higher mean intersection over union (mIoU) score of 0.8771 using only 0.98M parameters, demonstrating the effectiveness and potential practical value.</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

Lightweight metal surface defect segmentation method based on multi-scale feature fusion and knowledge distillation

  • 期刊论文

  • 发表日期：

    2025年5月25日

查看全文
  • Mingchun Li,
  • Yang Liu,
  • Dali Chen,
  • Xin Li

摘要

The shape, location, and size of surface defects on steel determine the quality of alloy products. Clearly, implementing automatic segmentation methods for surface defects is valuable. However, given the complexity of the surface defect segmentation task, achieving both segmentation accuracy and real-time prediction in intelligent methods is challenging, which limits their application prospects in the industry. To address the issue, we propose a lightweight deep learning model for defect segmentation tasks. It is a multi-scale feature fusion network (MFF-Metal) that captures effective feature representations through a feature fusion module and a multi-scale attention decoder. Moreover, a fusion loss function and soft label knowledge distillation are designed to improve performances while achieving lightweight prediction. In multiple experiments, compared with multiple existing deep learning models, our method achieves a higher mean intersection over union (mIoU) score of 0.8771 using only 0.98M parameters, demonstrating the effectiveness and potential practical value.