Scholarlink

<p>Fabric defect detection (FDD) is crucial in textile quality control, impacting yield and production costs. Traditional methods rely heavily on manual feature engineering, facing limitations due to complex textures. Recent deep learning approaches, particularly vision transformer (ViT)-based models, have shown promise but suffer from high computational complexity and parameter redundancy. This paper introduces the snake-shaped global–local cooperation feature shrinking pyramid network (SGL-FSPNet), a lightweight ViT-based architecture designed for efficient and precise FDD. SGL-FSPNet integrates a global–local interaction module (GLIM) to enhance feature discriminability and a snake-shaped feature shrinkage decoder (Snake-FSD) to improve cross-scale information aggregation. Experiments on the ZJU-Leaper dataset demonstrate superior performance, achieving peak pixel-level PPV and F1-scores of 89.69% and 88.21%, respectively, outperforming baseline models. This work establishes a novel lightweight architecture paradigm for industrial visual inspection. The code is available at <a href="https://github.com/chenyuntaot/SGL-FSPNet">https://github.com/chenyuntaot/SGL-FSPNet</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

Enhanced fabric defect detection via snake-shaped global–local feature shrinking network

  • 期刊论文

  • 发表日期：

    2026年4月6日

查看全文
  • Yuntao Chen,
  • Hao Liu,
  • Junjie Zhuang,
  • Jiuzhen Liang

摘要

Fabric defect detection (FDD) is crucial in textile quality control, impacting yield and production costs. Traditional methods rely heavily on manual feature engineering, facing limitations due to complex textures. Recent deep learning approaches, particularly vision transformer (ViT)-based models, have shown promise but suffer from high computational complexity and parameter redundancy. This paper introduces the snake-shaped global–local cooperation feature shrinking pyramid network (SGL-FSPNet), a lightweight ViT-based architecture designed for efficient and precise FDD. SGL-FSPNet integrates a global–local interaction module (GLIM) to enhance feature discriminability and a snake-shaped feature shrinkage decoder (Snake-FSD) to improve cross-scale information aggregation. Experiments on the ZJU-Leaper dataset demonstrate superior performance, achieving peak pixel-level PPV and F1-scores of 89.69% and 88.21%, respectively, outperforming baseline models. This work establishes a novel lightweight architecture paradigm for industrial visual inspection. The code is available at https://github.com/chenyuntaot/SGL-FSPNet.