Scholarlink

<p>The development of deep neural networks has significantly improved the accuracy of strip steel surface defect detection in industry. However, most existing models are based on high computing power. Restricted by the limited computing resources in high-speed industrial manufacturing environments, these models are difficult to achieve real-time and efficient detection in practice. To address these challenges, this paper proposes a detection network called Lightweight Detector with Dynamic-sample and Region-Attention Multi-Kernel Enhancement for Rapid Inspection (LDR-YOLO) to effectively improve computing speed while ensuring accuracy. Specifically, to significantly reduce the number of parameters and computation amount, a backbone network stacked with depthwise separable convolutions is introduced. The lightweight dynamic upsampling operator enhances the fusion effect of multi-scale features. In addition, the proposed regional attention multi-kernel module extracts multi-scale features through multi-scale convolution kernels. At the same time, it can capture local context information to ensure detection accuracy. Tests were conducted on both the NEU-DET and GC10-DET datasets. On the NEU-DET, our method uses only 1.64 million parameters to achieve 76.3 mAP@50<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> and 357 frames per second (FPS). This proves that the proposed method is suitable for real-time detection in resource-constrained industrial environments.</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

LDR-YOLO: lightweight dynamic upsampling and regional attention enhancement network for steel surface defect detection

  • 期刊论文

  • 发表日期：

    2025年12月27日

查看全文
  • Shanshan Tan,
  • Jinwen Tan,
  • Junfeng Man,
  • Junjie Ma,
  • Xiaoxue Ding

摘要

The development of deep neural networks has significantly improved the accuracy of strip steel surface defect detection in industry. However, most existing models are based on high computing power. Restricted by the limited computing resources in high-speed industrial manufacturing environments, these models are difficult to achieve real-time and efficient detection in practice. To address these challenges, this paper proposes a detection network called Lightweight Detector with Dynamic-sample and Region-Attention Multi-Kernel Enhancement for Rapid Inspection (LDR-YOLO) to effectively improve computing speed while ensuring accuracy. Specifically, to significantly reduce the number of parameters and computation amount, a backbone network stacked with depthwise separable convolutions is introduced. The lightweight dynamic upsampling operator enhances the fusion effect of multi-scale features. In addition, the proposed regional attention multi-kernel module extracts multi-scale features through multi-scale convolution kernels. At the same time, it can capture local context information to ensure detection accuracy. Tests were conducted on both the NEU-DET and GC10-DET datasets. On the NEU-DET, our method uses only 1.64 million parameters to achieve 76.3 mAP@50 \(\%\) % and 357 frames per second (FPS). This proves that the proposed method is suitable for real-time detection in resource-constrained industrial environments.