Scholarlink

<p>In recent years, ultra-high-definition (UHD) image processing has gained increasing attention due to the rapid advancements in imaging devices. Although benchmark datasets for UHD image dehazing have been proposed for the haze removal task, they remain limited by fixed image synthesis methods, restricting their ability to evaluate real-world complex hazy scenes. In this paper, we contribute the first real-world benchmark dataset for UHD image dehazing, called UHD-RealHaze. Specifically, we design a data acquisition system consisting of professional haze machines, enabling the collection of paired real hazy and corresponding haze-free UHD images. Moreover, to ensure dataset diversity, various haze patterns and scene perspectives are incorporated into our data collection process. Based on our proposed dataset, we retrain representative image dehazing algorithms and report both quantitative and qualitative benchmark results. We hope our study will further advance progress in this field. The proposed dataset will be publicly available (<a href="https://github.com/guanqiyuan/UHD-RealHaze">https://github.com/guanqiyuan/UHD-RealHaze</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

UHD-RealHaze: A Real-World Benchmark Dataset for Ultra-High-Definition Image Dehazing

  • 期刊论文

  • 发表日期：

    2025年7月4日

查看全文
  • Zongqi Huang,
  • Yiting Zhao,
  • Bolin Qian,
  • Ruizhe Wu,
  • Qiyuan Guan,
  • Xiang Chen

摘要

In recent years, ultra-high-definition (UHD) image processing has gained increasing attention due to the rapid advancements in imaging devices. Although benchmark datasets for UHD image dehazing have been proposed for the haze removal task, they remain limited by fixed image synthesis methods, restricting their ability to evaluate real-world complex hazy scenes. In this paper, we contribute the first real-world benchmark dataset for UHD image dehazing, called UHD-RealHaze. Specifically, we design a data acquisition system consisting of professional haze machines, enabling the collection of paired real hazy and corresponding haze-free UHD images. Moreover, to ensure dataset diversity, various haze patterns and scene perspectives are incorporated into our data collection process. Based on our proposed dataset, we retrain representative image dehazing algorithms and report both quantitative and qualitative benchmark results. We hope our study will further advance progress in this field. The proposed dataset will be publicly available (https://github.com/guanqiyuan/UHD-RealHaze).