Scholarlink

<p>Surface defect detection is crucial for industrial quality control, yet existing lightweight networks struggle to balance efficiency, global context, and segmentation accuracy for small defects. To address this issue, this study proposes DGFNet, a lightweight defect segmentation model integrating the Partial Deep Dynamic Attention (PDA) and Grouped Multiscale Feature Extraction Block (GMB). PDA adaptively fits defect boundaries via feature-space offsets, while GMB employs grouped convolutions and cross-channel fusion to capture both local and global patterns. A Flow-Guided Upsampling strategy refines edge reconstruction, and Cross-Hierarchical Supervision enhances sub-pixel sensitivity. DGFNet achieves an exceptional trade-off between accuracy and computational cost through a holistic lightweight design-combining a compact backbone, separable convolutions, and GMB’s grouped scheme. It attains 83.05% mIoU on NEU-Seg dataset, outperforming Fast-SCNN and SegFormer by 6.66% and 2.29%, respectively, while maintaining only 1.725 GFLOPs. State-of-the-art performance is also achieved on SMCD (93.72%) and Cracks (76.08%) datasets, validating its efficiency and effectiveness for fine-grained defect segmentation.</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

Dgfnet: deformable attention and flow-guided feature learning for fine-grained surface defect segmentation

  • 期刊论文

  • 发表日期：

    2025年11月24日

查看全文
  • Shengyi Chen,
  • Xiufang Zhu,
  • Jiqiang Wang,
  • Xinmin Chen,
  • Feifan Yu,
  • Sichen Li,
  • Shanhong Ye

摘要

Surface defect detection is crucial for industrial quality control, yet existing lightweight networks struggle to balance efficiency, global context, and segmentation accuracy for small defects. To address this issue, this study proposes DGFNet, a lightweight defect segmentation model integrating the Partial Deep Dynamic Attention (PDA) and Grouped Multiscale Feature Extraction Block (GMB). PDA adaptively fits defect boundaries via feature-space offsets, while GMB employs grouped convolutions and cross-channel fusion to capture both local and global patterns. A Flow-Guided Upsampling strategy refines edge reconstruction, and Cross-Hierarchical Supervision enhances sub-pixel sensitivity. DGFNet achieves an exceptional trade-off between accuracy and computational cost through a holistic lightweight design-combining a compact backbone, separable convolutions, and GMB’s grouped scheme. It attains 83.05% mIoU on NEU-Seg dataset, outperforming Fast-SCNN and SegFormer by 6.66% and 2.29%, respectively, while maintaining only 1.725 GFLOPs. State-of-the-art performance is also achieved on SMCD (93.72%) and Cracks (76.08%) datasets, validating its efficiency and effectiveness for fine-grained defect segmentation.