Scholarlink

<p>Robust and real-time sea-sky line detection is crucial for maritime applications, yet challenging due to environmental complexities and edge device constraints. This paper introduces a novel framework that integrates a lightweight semantic segmentation network (DCEUnet) with a dual multi-scale fusion mechanism to achieve sub-pixel accuracy. DCEUnet, with less than 0.4M parameters, provides coarse segmentation, defining a dynamic Region of Interest (ROI). Within this ROI, a refinement module fuses multi-scale edges with a learned spatial confidence prior. Extensive evaluations on MU-SID, SMD, BD, and TMD datasets demonstrate state-of-the-art performance, with the segmentation core achieving 98.18% IoU and the complete system operating at 20–50 FPS with median position errors of 0.86–1.87 pixels and median angular errors of 0.13–0.91 degrees. The source code and sample data are available at: <a href="https://github.com/kaibaJC/ESSLD">https://github.com/kaibaJC/ESSLD</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

Efficient sea-sky line detection via lightweight segmentation and dual multi-scale fusion

  • 期刊论文

  • 发表日期：

    2026年4月29日

查看全文
  • Jialuo Chen,
  • Zhiwu Hu

摘要

Robust and real-time sea-sky line detection is crucial for maritime applications, yet challenging due to environmental complexities and edge device constraints. This paper introduces a novel framework that integrates a lightweight semantic segmentation network (DCEUnet) with a dual multi-scale fusion mechanism to achieve sub-pixel accuracy. DCEUnet, with less than 0.4M parameters, provides coarse segmentation, defining a dynamic Region of Interest (ROI). Within this ROI, a refinement module fuses multi-scale edges with a learned spatial confidence prior. Extensive evaluations on MU-SID, SMD, BD, and TMD datasets demonstrate state-of-the-art performance, with the segmentation core achieving 98.18% IoU and the complete system operating at 20–50 FPS with median position errors of 0.86–1.87 pixels and median angular errors of 0.13–0.91 degrees. The source code and sample data are available at: https://github.com/kaibaJC/ESSLD.