Scholarlink

<p>Ensuring the operational safety and integrity of gas pipelines, which are critical components of urban infrastructure, is paramount. However, the accurate and efficient detection of defects remains a significant challenge. To address this challenge, we propose a Lite Distance-Content aware DEtection TRansformer (L-DC-DETR). Our main contributions are threefold: First, to reduce computational complexity, we introduce a lite encoder that reduces computational complexity by alternately updating feature maps at different scales and selectively processing only the top-K high-resolution features. Second, to improve the quality of decoder queries, we introduce a Cascade High-Quality Query Selection (CHQQS) module that evaluates query quality, retaining high-quality queries as decoder input and mitigating interference from low-quality ones. Finally, to enhance the detection of small defects, we design a Distance-Content aware Deformable Attention (DCDA) decoder, which improves focus on small targets by calculating the distance and content correlation between features. Experimental results on our self-built gas pipeline defect (PIP-DET) dataset and the public NEU-DET dataset demonstrate that L-DC-DETR outperforms existing mainstream methods, verifying its effectiveness.The source code is available at <a href="https://github.com/delerpo/LDC-DETR">https://github.com/delerpo/LDC-DETR</a>.</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

L-DC-DETR: A Lite Distance-Content Aware DETR for Gas Pipeline Defect Detection

  • 期刊论文

  • 发表日期：

    2026年1月14日

查看全文
  • Chaofan Chen,
  • Ting Zhang,
  • Cong Ma,
  • Zhaoying Liu,
  • Sadaqat ur Rehman,
  • Yanan Shi,
  • Amr Munshi

摘要

Ensuring the operational safety and integrity of gas pipelines, which are critical components of urban infrastructure, is paramount. However, the accurate and efficient detection of defects remains a significant challenge. To address this challenge, we propose a Lite Distance-Content aware DEtection TRansformer (L-DC-DETR). Our main contributions are threefold: First, to reduce computational complexity, we introduce a lite encoder that reduces computational complexity by alternately updating feature maps at different scales and selectively processing only the top-K high-resolution features. Second, to improve the quality of decoder queries, we introduce a Cascade High-Quality Query Selection (CHQQS) module that evaluates query quality, retaining high-quality queries as decoder input and mitigating interference from low-quality ones. Finally, to enhance the detection of small defects, we design a Distance-Content aware Deformable Attention (DCDA) decoder, which improves focus on small targets by calculating the distance and content correlation between features. Experimental results on our self-built gas pipeline defect (PIP-DET) dataset and the public NEU-DET dataset demonstrate that L-DC-DETR outperforms existing mainstream methods, verifying its effectiveness.The source code is available at https://github.com/delerpo/LDC-DETR.