Scholarlink

<p>Rapid and accurate detection of surface defect for strip steel is of utmost importance in ensuring the quality of industrial manufacturing. However, some commonly used deep learning-based defect detection methods, which prioritize high performance, are unsuitable for resource-constrained embedded instruments due to their intricate network models and computational requirements. This paper proposes a lightweight defect detection model based on NanoDet-Plus, achieving good performance in terms of precision and timeliness. First, a Ghost block is introduced into the feature pyramid to strengthen feature extraction. Second, a spatial and channel attention (SCA) module is employed to effectively merge feature maps of different scales. Finally, the multi-scale detection head is adopted to improve the detection ability of defects with varying scales. Experimental results show that the proposed network has only 1.17&#xa0;MB in size, but achieves the mean average precision (mAP<sub>@0.5:0.95</sub>) of 46.02% and 34.14% on two public datasets, outperforming other state-of-the-art lightweight models. In addition, the proposed network can run on Raspberry Pi 4B at a detection speed of 11 frames per second (FPS), indicating its potential deployment on front-end devices with limited computing power for real-time surface defect detection.</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

Steel surface defect detection based on lightweight network with attention feature fusion and multi-scale detection head

  • 期刊论文

  • 发表日期：

    2026年1月6日

查看全文
  • Yunke Fu,
  • Yong Wang,
  • Zhuo Cheng,
  • Yingyu Li

摘要

Rapid and accurate detection of surface defect for strip steel is of utmost importance in ensuring the quality of industrial manufacturing. However, some commonly used deep learning-based defect detection methods, which prioritize high performance, are unsuitable for resource-constrained embedded instruments due to their intricate network models and computational requirements. This paper proposes a lightweight defect detection model based on NanoDet-Plus, achieving good performance in terms of precision and timeliness. First, a Ghost block is introduced into the feature pyramid to strengthen feature extraction. Second, a spatial and channel attention (SCA) module is employed to effectively merge feature maps of different scales. Finally, the multi-scale detection head is adopted to improve the detection ability of defects with varying scales. Experimental results show that the proposed network has only 1.17 MB in size, but achieves the mean average precision (mAP@0.5:0.95) of 46.02% and 34.14% on two public datasets, outperforming other state-of-the-art lightweight models. In addition, the proposed network can run on Raspberry Pi 4B at a detection speed of 11 frames per second (FPS), indicating its potential deployment on front-end devices with limited computing power for real-time surface defect detection.