Scholarlink

<p>Steel surface defect detection faces challenges such as tiny-target perception, heavy texture interference, and semantic inconsistency. Existing methods often show limited shallow-layer modeling, weak multi-scale fusion, and supervision conflicts. To address these issues, we propose FreqAlignNet, a frequency-aligned, three-stage defect detection framework that enhances accuracy and robustness through joint spatial–frequency modeling and structural decoupling. FreqAlignNet comprises: (1) AHGFNet, which strengthens early small-defect boundary and frequency feature perception; (2) FSCLFN, which enforces semantic consistency and cross-scale contextual alignment; and (3) DSENet, a dual-branch design that decouples classification and regression while improving boundary sensitivity for small targets. Experiments on NEU-DET, GC10-DET, and Severstal show that FreqAlignNet surpasses state-of-the-art methods in mAP, parameter efficiency, and inference speed. It also demonstrates strong robustness in detecting defects with blurred boundaries, providing an efficient and deployable solution for industrial applications.</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

FreqAlignNet: bridging local detail and global semantics via frequency-aligned representation for steel defect detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月1日

查看全文
  • Yuzhen Zhao,
  • Xun Li,
  • Yang Zhao,
  • Zhun Guo,
  • Jianjing Gao,
  • Ruijuan Yao,
  • Baoxi Yuan

摘要

Steel surface defect detection faces challenges such as tiny-target perception, heavy texture interference, and semantic inconsistency. Existing methods often show limited shallow-layer modeling, weak multi-scale fusion, and supervision conflicts. To address these issues, we propose FreqAlignNet, a frequency-aligned, three-stage defect detection framework that enhances accuracy and robustness through joint spatial–frequency modeling and structural decoupling. FreqAlignNet comprises: (1) AHGFNet, which strengthens early small-defect boundary and frequency feature perception; (2) FSCLFN, which enforces semantic consistency and cross-scale contextual alignment; and (3) DSENet, a dual-branch design that decouples classification and regression while improving boundary sensitivity for small targets. Experiments on NEU-DET, GC10-DET, and Severstal show that FreqAlignNet surpasses state-of-the-art methods in mAP, parameter efficiency, and inference speed. It also demonstrates strong robustness in detecting defects with blurred boundaries, providing an efficient and deployable solution for industrial applications.