Scholarlink

<p>Sonar images, characterized by low resolution and high noise, present significant challenges for effective segmentation. Existing networks struggle to suppress the inherent noise in these images. We propose SonarRanDiff, a novel segmentation network tailored for sonar images. For the first time, we introduce a diffusion model to sonar image segmentation, incorporating a random fusion module, RanFusion, which boosts convergence speed and improves accuracy by effectively integrating original image and mask information. Comprehensive experiments show that SonarRanDiff surpasses state-of-the-art methods by at least 4.5% on our sonar dataset, underscoring its superior ability to address the unique challenges of sonar imaging. The code is available at: <a href="https://github.com/mzh529/SonarRanDiff.git">https://github.com/mzh529/SonarRanDiff.git</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

Enhancing sonar image segmentation with random fusion in a diffusion model framework

  • 期刊论文

  • 发表日期：

    2025年4月1日

查看全文
  • Zhihao Ma,
  • Weiliang Meng,
  • Xixi Zhao,
  • Longyu Jiang

摘要

Sonar images, characterized by low resolution and high noise, present significant challenges for effective segmentation. Existing networks struggle to suppress the inherent noise in these images. We propose SonarRanDiff, a novel segmentation network tailored for sonar images. For the first time, we introduce a diffusion model to sonar image segmentation, incorporating a random fusion module, RanFusion, which boosts convergence speed and improves accuracy by effectively integrating original image and mask information. Comprehensive experiments show that SonarRanDiff surpasses state-of-the-art methods by at least 4.5% on our sonar dataset, underscoring its superior ability to address the unique challenges of sonar imaging. The code is available at: https://github.com/mzh529/SonarRanDiff.git.