Scholarlink

<p>Accurate detection of small defects is critical for intelligent manufacturing quality control, as undetected flaws may lead to serious product quality risks. However, achieving reliable detection remains highly challenging in industrial environments, particularly under complex background interference where conventional methods often suffer from feature loss and misidentification. To address these issues, this paper proposes the Visual Defect Detection Network (VisDefectNet), a lightweight and fine-grained defect detection network built upon YOLOv11s and tailored for small object detection. The proposed framework incorporates three synergistic innovations: a Spatial-Shift Channel Mixer Module (SS-CMM) that performs parameter-free spatial-channel recombination at shallow layers to preserve microstructural features, a Cross Stage Partial with Top-k Sparse Attention mechanism (C2PSA-Topk) that suppresses background noise via sparsified attention regions, and a Multi-Scale Kernel Space Upsampling (MS-KSU) module that restores defect details through content-aware offset prediction. Furthermore, integrating Adaptive Threshold Focal Loss (ATFLoss) enhances detection performance across varying defect scales. Extensive experiments demonstrate that the proposed VisDefectNet consistently outperforms existing methods in both accuracy and efficiency, achieving mAP@50 scores of 78.4%, 95.7%, and 85.9% on the NEU-DET, PCB-DET, and DsPCBSD+ datasets, respectively, with only 8.3M parameters and an inference speed of 79 FPS. These results establish VisDefectNet as a high-precision and efficient solution for practical industrial small-defect inspection, offering significant value for automated quality control systems.</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

VisDefectNet: A lightweight and fine-grained detection network for small defects in complex industrial scenes

  • 期刊论文

  • 发表日期：

    2026年4月15日

查看全文
  • Qingling Xia,
  • Dinghao Luo,
  • Hong Zheng,
  • Gen Li,
  • Yan Li,
  • Yun Zhao,
  • Lingxiao Li,
  • Bin Jiang

摘要

Accurate detection of small defects is critical for intelligent manufacturing quality control, as undetected flaws may lead to serious product quality risks. However, achieving reliable detection remains highly challenging in industrial environments, particularly under complex background interference where conventional methods often suffer from feature loss and misidentification. To address these issues, this paper proposes the Visual Defect Detection Network (VisDefectNet), a lightweight and fine-grained defect detection network built upon YOLOv11s and tailored for small object detection. The proposed framework incorporates three synergistic innovations: a Spatial-Shift Channel Mixer Module (SS-CMM) that performs parameter-free spatial-channel recombination at shallow layers to preserve microstructural features, a Cross Stage Partial with Top-k Sparse Attention mechanism (C2PSA-Topk) that suppresses background noise via sparsified attention regions, and a Multi-Scale Kernel Space Upsampling (MS-KSU) module that restores defect details through content-aware offset prediction. Furthermore, integrating Adaptive Threshold Focal Loss (ATFLoss) enhances detection performance across varying defect scales. Extensive experiments demonstrate that the proposed VisDefectNet consistently outperforms existing methods in both accuracy and efficiency, achieving mAP@50 scores of 78.4%, 95.7%, and 85.9% on the NEU-DET, PCB-DET, and DsPCBSD+ datasets, respectively, with only 8.3M parameters and an inference speed of 79 FPS. These results establish VisDefectNet as a high-precision and efficient solution for practical industrial small-defect inspection, offering significant value for automated quality control systems.