Scholarlink

<p>Surface defect detection is crucial for ensuring industrial product quality. However, considering the intricate texture structures and noise of these products, existing object detection networks have struggled to achieve both accurate and rapid detection while maintaining low complexity. This paper proposed LEIS-Net, a lightweight and efficient industrial surface defect detection network. First, to reduce backbone network complexity while optimizing downsampling and feature extraction, an edge information extraction module (EIEM) was designed for the front end of the network. Additionally, a multi-scale downsampling (MSD) module based on Haar wavelets, and a lightweight feature enhancement module (LFEM) were integrated into the middle of the network. Second, to improve neck’s efficiency in fusing multi-scale features, channel alignment and the bi-directional feature pyramid network (BiFPN) fusion were adopted in the network. Finally, to enhance efficiency and generalization, the network employs shared convolutions, reparameterization, and adaptive scaling. These techniques reduce complexity while maintaining performance. The experimental results on the NEU-DET and PVEL-AD datasets showed that the proposed LEIS-Net outperformed existing methods in balancing accuracy, speed, and network complexity, demonstrating its effectiveness and superiority in industrial 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

LEIS-net: lightweight and efficient industrial surface defect detection network

  • 期刊论文

  • 发表日期：

    2025年8月6日

查看全文
  • Chunlong Yang,
  • Donghao Lv,
  • Xu Tian,
  • Chenzhi Wang,
  • Peihong Yang,
  • Yong Zhang

摘要

Surface defect detection is crucial for ensuring industrial product quality. However, considering the intricate texture structures and noise of these products, existing object detection networks have struggled to achieve both accurate and rapid detection while maintaining low complexity. This paper proposed LEIS-Net, a lightweight and efficient industrial surface defect detection network. First, to reduce backbone network complexity while optimizing downsampling and feature extraction, an edge information extraction module (EIEM) was designed for the front end of the network. Additionally, a multi-scale downsampling (MSD) module based on Haar wavelets, and a lightweight feature enhancement module (LFEM) were integrated into the middle of the network. Second, to improve neck’s efficiency in fusing multi-scale features, channel alignment and the bi-directional feature pyramid network (BiFPN) fusion were adopted in the network. Finally, to enhance efficiency and generalization, the network employs shared convolutions, reparameterization, and adaptive scaling. These techniques reduce complexity while maintaining performance. The experimental results on the NEU-DET and PVEL-AD datasets showed that the proposed LEIS-Net outperformed existing methods in balancing accuracy, speed, and network complexity, demonstrating its effectiveness and superiority in industrial surface defect detection.