Scholarlink

<p>In the field of industrial defect detection, achieving precise small-target detection under high-speed conditions remains a key challenge in computer vision, particularly in time-sensitive scenarios such as special-material production line inspection. To address this problem, a lightweight high-speed DETR model (LHS-DETR) for micro-defects detection is proposed. The model first employs Faster Space-Frequency Selective Network (FSFSNet). By utilizing a fractional-order Gabor transform, it effectively distinguishes defect regions from normal backgrounds, enhancing real-time detection performance while lowering both computational cost and model parameters. Next, a linear efficient attention fusion (LEAF) module is incorporated into the encoder, lowering computational complexity from quadratic to linear order and enhancing the interaction between local features and global context, which further accelerates inference. Finally, a multi-scale star fusion core(MSSFC) is integrated into the model, enabling dynamic aggregation and semantic enhancement of multi-scale features through parallel branches and gated product operations. LHS-DETR demonstrates superior performance against existing SOTA methods on two industrial production line datasets, achieving mAP@50 values of 78.4% and 79.4% respectively, while maintaining a compact parameter footprint of 13.7M and low computational complexity of 42.3 GFLOPs.</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

LHS-DETR: a lightweight high-speed DETR model for micro-defects detection in industrial specialty materials

  • 期刊论文

  • 发表日期：

    2026年3月17日

查看全文
  • Chentao Gong,
  • Kui Qian,
  • Yue Deng,
  • Lei Yang

摘要

In the field of industrial defect detection, achieving precise small-target detection under high-speed conditions remains a key challenge in computer vision, particularly in time-sensitive scenarios such as special-material production line inspection. To address this problem, a lightweight high-speed DETR model (LHS-DETR) for micro-defects detection is proposed. The model first employs Faster Space-Frequency Selective Network (FSFSNet). By utilizing a fractional-order Gabor transform, it effectively distinguishes defect regions from normal backgrounds, enhancing real-time detection performance while lowering both computational cost and model parameters. Next, a linear efficient attention fusion (LEAF) module is incorporated into the encoder, lowering computational complexity from quadratic to linear order and enhancing the interaction between local features and global context, which further accelerates inference. Finally, a multi-scale star fusion core(MSSFC) is integrated into the model, enabling dynamic aggregation and semantic enhancement of multi-scale features through parallel branches and gated product operations. LHS-DETR demonstrates superior performance against existing SOTA methods on two industrial production line datasets, achieving mAP@50 values of 78.4% and 79.4% respectively, while maintaining a compact parameter footprint of 13.7M and low computational complexity of 42.3 GFLOPs.