Scholarlink

<p>Autonomous marine navigation systems increasingly rely on accurate water surface boundary detection for safe obstacle avoidance and path planning. However, deploying segmentation models on resource-constrained unmanned surface vehicles (USVs) poses significant challenges: Existing methods either achieve high accuracy with prohibitive computational costs unsuitable for edge devices, or sacrifice boundary precision for efficiency. This paper presents EdgeFlowNet, a lightweight and environment-robust edge-aware network specifically designed for real-time water surface segmentation. The proposed architecture features three key innovations: (1) a dynamic importance weighting (DIW) module that spatially adapts the fusion of edge and semantic features based on local water surface characteristics, (2) a hybrid dual-branch design integrating classical edge operators (Sobel, Laplacian) with learnable convolutions for enhanced boundary localization, and (3) extreme parameter efficiency through binary classification optimization achieving only 32.6 to 81.8&#xa0;K parameters across three configurations. Comprehensive experiments on three maritime datasets (MODD2, MaSTr1325, and WaterScenes) demonstrate that EdgeFlowNet-Standard (81.8&#xa0;K parameters) achieves 99.02% mIoU on MODD2 and exhibits superior cross-dataset generalization, outperforming PiDiNet on MaSTr1325 (+0.13%) and WaterScenes (+0.52%) while using only 11.5% of its parameters. On edge devices (Axera AX650x), EdgeFlowNet achieves real-time performance of 67.7 FPS with only 2.6W power consumption, enabling practical deployment for USV navigation.</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

EdgeFlowNet: lightweight edge-aware network for high-precision water surface segmentation

  • 期刊论文

  • 发表日期：

    2026年5月7日

查看全文
  • Xiao-ting Guo,
  • Yu-xuan Liao,
  • Hong-liang Wang,
  • Shuai-kang Sun

摘要

Autonomous marine navigation systems increasingly rely on accurate water surface boundary detection for safe obstacle avoidance and path planning. However, deploying segmentation models on resource-constrained unmanned surface vehicles (USVs) poses significant challenges: Existing methods either achieve high accuracy with prohibitive computational costs unsuitable for edge devices, or sacrifice boundary precision for efficiency. This paper presents EdgeFlowNet, a lightweight and environment-robust edge-aware network specifically designed for real-time water surface segmentation. The proposed architecture features three key innovations: (1) a dynamic importance weighting (DIW) module that spatially adapts the fusion of edge and semantic features based on local water surface characteristics, (2) a hybrid dual-branch design integrating classical edge operators (Sobel, Laplacian) with learnable convolutions for enhanced boundary localization, and (3) extreme parameter efficiency through binary classification optimization achieving only 32.6 to 81.8 K parameters across three configurations. Comprehensive experiments on three maritime datasets (MODD2, MaSTr1325, and WaterScenes) demonstrate that EdgeFlowNet-Standard (81.8 K parameters) achieves 99.02% mIoU on MODD2 and exhibits superior cross-dataset generalization, outperforming PiDiNet on MaSTr1325 (+0.13%) and WaterScenes (+0.52%) while using only 11.5% of its parameters. On edge devices (Axera AX650x), EdgeFlowNet achieves real-time performance of 67.7 FPS with only 2.6W power consumption, enabling practical deployment for USV navigation.