Scholarlink

<p>The scarcity of high-quality datasets severely constrains research in sonar image analysis. Existing image generation methods often produce low-fidelity results with sparse details due to uncontrollable randomness. To address these limitations, we introduce the Hierarchical Segmentation-guided Diffusion Model (HSD-Sonar), a novel framework designed for high-fidelity sonar image synthesis. HSD-Sonar employs a hierarchical optimization strategy: at the image level, it incorporates an adaptive mechanism to adjust diffusion steps based on input characteristics and utilizes segmentation masks to precisely regulate target regions, and integrates a contrast enhancement strategy to sharpen fine-grained details and semantic consistency. At the feature level, it augments spatial information with mask data and optimizes the objective function via a region importance-based SG-Loss, effectively mitigating background noise and improving perceptual realism. Extensive experiments on multiple public sonar segmentation datasets demonstrate that HSD-Sonar significantly outperforms existing methods in generating high-quality, semantically consistent sonar images. Our research offers a robust and flexible solution, facilitating advancements in sonar imagery interpretation. The code is available at <a href="https://github.com/mzh529/HSD-Sonar">https://github.com/mzh529/HSD-Sonar</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

Hierarchical segmentation-guided diffusion framework for high-fidelity sonar image generation

  • 期刊论文

  • 发表日期：

    2026年5月6日

查看全文
  • Zhihao Ma,
  • Weiliang Meng,
  • Chenghanxue Tang,
  • Longyu Jiang

摘要

The scarcity of high-quality datasets severely constrains research in sonar image analysis. Existing image generation methods often produce low-fidelity results with sparse details due to uncontrollable randomness. To address these limitations, we introduce the Hierarchical Segmentation-guided Diffusion Model (HSD-Sonar), a novel framework designed for high-fidelity sonar image synthesis. HSD-Sonar employs a hierarchical optimization strategy: at the image level, it incorporates an adaptive mechanism to adjust diffusion steps based on input characteristics and utilizes segmentation masks to precisely regulate target regions, and integrates a contrast enhancement strategy to sharpen fine-grained details and semantic consistency. At the feature level, it augments spatial information with mask data and optimizes the objective function via a region importance-based SG-Loss, effectively mitigating background noise and improving perceptual realism. Extensive experiments on multiple public sonar segmentation datasets demonstrate that HSD-Sonar significantly outperforms existing methods in generating high-quality, semantically consistent sonar images. Our research offers a robust and flexible solution, facilitating advancements in sonar imagery interpretation. The code is available at https://github.com/mzh529/HSD-Sonar.