Scholarlink

<p>Defect detection in sawn lumber is crucial for production quality. However, the similarity in texture and color among wood defects, along with unclear boundaries and wide variations in size and shape, poses significant challenges. To address these issues, this paper proposes a novel network based on enhanced features and content-aware fusion (ECF-Net). First, the Transformer architecture is employed to extract multi-scale features of the defects. Then, a 3D attention mechanism is introduced to refine the details of the multi-scale features in both semantic and spatial dimensions, ensuring pixel-level detection accuracy. Next, a detail-guided feature enhancement module is designed, using low-level features to guide high-level ones, helping the latter better capture the structural information from the former. Finally, a content-aware fusion module is developed to optimize the attention weights from coarse to fine using the enhanced features, assisting the segmentation head in better recognizing different defects. Experimental results on the rubber and pine datasets demonstrate that the proposed ECF-Net achieves mIoU scores of 74.20% and 83.21%, respectively, indicating competitive performance. Extensive ablation studies further confirm the effectiveness of each component. Code will be available on <a href="https://github.com/hhuzzz/ECF-Net">GitHub</a>.</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

ECF-Net: lumber defect segmentation network with enhanced feature and content-aware fusion

  • 期刊论文

  • 发表日期：

    2025年4月29日

查看全文
  • Huan Hu,
  • Fengwen Liu,
  • Nan Su,
  • Wenqiang Hu

摘要

Defect detection in sawn lumber is crucial for production quality. However, the similarity in texture and color among wood defects, along with unclear boundaries and wide variations in size and shape, poses significant challenges. To address these issues, this paper proposes a novel network based on enhanced features and content-aware fusion (ECF-Net). First, the Transformer architecture is employed to extract multi-scale features of the defects. Then, a 3D attention mechanism is introduced to refine the details of the multi-scale features in both semantic and spatial dimensions, ensuring pixel-level detection accuracy. Next, a detail-guided feature enhancement module is designed, using low-level features to guide high-level ones, helping the latter better capture the structural information from the former. Finally, a content-aware fusion module is developed to optimize the attention weights from coarse to fine using the enhanced features, assisting the segmentation head in better recognizing different defects. Experimental results on the rubber and pine datasets demonstrate that the proposed ECF-Net achieves mIoU scores of 74.20% and 83.21%, respectively, indicating competitive performance. Extensive ablation studies further confirm the effectiveness of each component. Code will be available on GitHub.