Scholarlink

<p>Sewer pipeline defect detection is critical for urban infrastructure maintenance, but faces challenges like occlusions, uneven lighting, chromatic homogeneity, and lack of geometric texture. Conventional single-modal RGB-based methods struggle in these complex environments due to their inherent lack of 3D structural perception. To address this, we propose the <b>Depth-Guided Multimodal Attention Fusion Network (DGMA-Net)</b>, a novel framework built upon YOLOv8n. DGMA-Net innovatively integrates depth information estimated from RGB images with visual features through a cross-modal attention architecture. This architecture features a Lightweight Cross-Attention (LCA) module for efficient cross-modal geometric reasoning and a Spatial-Channel Coordinated Fusion (SCCF) module for dynamic spatial weighting and feature refinement. A learnable weighted adaptive mechanism further optimizes the fusion. In evaluations on our SewerDefect-3K dataset and public dataset, DGMA-Net demonstrates significant improvements in accuracy and robustness compared to leading single-modal RGB detectors. Experiments show that depth-guided multimodal fusion enhances defect detection performance in challenging sewer 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

DGMA-Net: depth-guided multimodal attention fusion network for sewer pipe defect detection on inspection robots

  • 期刊论文

  • 发表日期：

    2026年2月17日

查看全文
  • Shiliang Pan,
  • Ying Zhang,
  • Wenjing Zhuang,
  • Xiangcheng Jiang,
  • Yongjun Xu,
  • Kai Zhao,
  • Qianlong Feng

摘要

Sewer pipeline defect detection is critical for urban infrastructure maintenance, but faces challenges like occlusions, uneven lighting, chromatic homogeneity, and lack of geometric texture. Conventional single-modal RGB-based methods struggle in these complex environments due to their inherent lack of 3D structural perception. To address this, we propose the Depth-Guided Multimodal Attention Fusion Network (DGMA-Net), a novel framework built upon YOLOv8n. DGMA-Net innovatively integrates depth information estimated from RGB images with visual features through a cross-modal attention architecture. This architecture features a Lightweight Cross-Attention (LCA) module for efficient cross-modal geometric reasoning and a Spatial-Channel Coordinated Fusion (SCCF) module for dynamic spatial weighting and feature refinement. A learnable weighted adaptive mechanism further optimizes the fusion. In evaluations on our SewerDefect-3K dataset and public dataset, DGMA-Net demonstrates significant improvements in accuracy and robustness compared to leading single-modal RGB detectors. Experiments show that depth-guided multimodal fusion enhances defect detection performance in challenging sewer environments.