Scholarlink

<p>To address the precision-efficiency trade-off in detecting small fabric defects under strict computational constraints, we propose OHS-Net, an edge-oriented lightweight hybrid architecture centered on a novel Feature Purification paradigm. It intrinsically unifies efficient local CNN extraction with refined global Transformer reasoning. To tackle complex, periodic fabric backgrounds, we introduce an Orthogonal Channel Attention (OCA) mechanism. By enforcing mathematical orthogonality, OCA decorrelates feature channels, suppressing high-frequency noise and enhancing micro-defect discriminability with negligible computational burden. Furthermore, to prevent losing critical micro-defect signatures during feature encoding, we employ Haar Wavelet Downsampling (HWD), substituting lossy spatial pooling with information-preserving frequency-domain downsampling. To adapt this purified feature space for rigorous industrial deployment requirements, we integrate a Semantics and Detail Infusion (SDI) module alongside Focal-CIoU loss and channel-wise knowledge distillation. These optimizations resolve extreme scale imbalances and boost hard-sample regression precision with zero additional inference cost. Experiments demonstrate OHS-Net outperforms the YOLO-DETR baseline by absolute margins of 2.8% (mAP@0.5) and 1.5% (mAP@0.5:0.95), while reducing parameters by 5.09% and GFLOPs by 5.98% at 220.6 FPS. This optimally balanced architecture demonstrates high theoretical viability and strong potential for robust industrial edge deployment.</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

A lightweight orthogonal hybrid network with feature purification for real-time industrial fabric defect inspection

  • 期刊论文

  • 发表日期：

    2026年4月28日

查看全文
  • Ming Li,
  • Jie Hu,
  • Zhiqi He,
  • Shengli Zhu

摘要

To address the precision-efficiency trade-off in detecting small fabric defects under strict computational constraints, we propose OHS-Net, an edge-oriented lightweight hybrid architecture centered on a novel Feature Purification paradigm. It intrinsically unifies efficient local CNN extraction with refined global Transformer reasoning. To tackle complex, periodic fabric backgrounds, we introduce an Orthogonal Channel Attention (OCA) mechanism. By enforcing mathematical orthogonality, OCA decorrelates feature channels, suppressing high-frequency noise and enhancing micro-defect discriminability with negligible computational burden. Furthermore, to prevent losing critical micro-defect signatures during feature encoding, we employ Haar Wavelet Downsampling (HWD), substituting lossy spatial pooling with information-preserving frequency-domain downsampling. To adapt this purified feature space for rigorous industrial deployment requirements, we integrate a Semantics and Detail Infusion (SDI) module alongside Focal-CIoU loss and channel-wise knowledge distillation. These optimizations resolve extreme scale imbalances and boost hard-sample regression precision with zero additional inference cost. Experiments demonstrate OHS-Net outperforms the YOLO-DETR baseline by absolute margins of 2.8% (mAP@0.5) and 1.5% (mAP@0.5:0.95), while reducing parameters by 5.09% and GFLOPs by 5.98% at 220.6 FPS. This optimally balanced architecture demonstrates high theoretical viability and strong potential for robust industrial edge deployment.