Scholarlink

<p>Low-light image enhancement aims to recover visually natural and detail-preserving images from scenes captured under insufficient illumination. However, existing methods often improve brightness at the cost of noise amplification, structural degradation, and locally inconsistent textures, making it difficult to simultaneously preserve dark-region details, edge fidelity, and color stability. To address these issues, we propose a Luminance Residual Diffusion Network (LumiRD), a two-stage framework that combines degradation-aware representation refinement with residual-space restoration. Specifically, an Illumination Manifold Unfolding Module (IMUM) jointly models spatial, gradient, and frequency-domain cues to unfold suppressed low-light responses and enhance the structural discriminability of dark regions. Building on this, a Residual Diffusion Module (RDM) performs residual-guided refinement from a coarse enhancement result by integrating global exposure correction with local detail compensation, thereby reducing luminance accumulation, texture drift, and over-enhancement during restoration. Experimental results on LOL, SICE-Part2, and LSRW demonstrate that LumiRD achieves competitive quantitative results and favorable visual quality under diverse illumination conditions. Our code is publicly available at <a href="https://github.com/ShengyuFang/lumird.git">https://github.com/ShengyuFang/lumird.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

Progressive luminance residual diffusion with manifold-based representation refinement for low-light image enhancement

  • 期刊论文

  • 发表日期：

    2026年5月9日

查看全文
  • Shengyu Fang,
  • Kai Zhang,
  • Li Liu,
  • Guihai Zhao

摘要

Low-light image enhancement aims to recover visually natural and detail-preserving images from scenes captured under insufficient illumination. However, existing methods often improve brightness at the cost of noise amplification, structural degradation, and locally inconsistent textures, making it difficult to simultaneously preserve dark-region details, edge fidelity, and color stability. To address these issues, we propose a Luminance Residual Diffusion Network (LumiRD), a two-stage framework that combines degradation-aware representation refinement with residual-space restoration. Specifically, an Illumination Manifold Unfolding Module (IMUM) jointly models spatial, gradient, and frequency-domain cues to unfold suppressed low-light responses and enhance the structural discriminability of dark regions. Building on this, a Residual Diffusion Module (RDM) performs residual-guided refinement from a coarse enhancement result by integrating global exposure correction with local detail compensation, thereby reducing luminance accumulation, texture drift, and over-enhancement during restoration. Experimental results on LOL, SICE-Part2, and LSRW demonstrate that LumiRD achieves competitive quantitative results and favorable visual quality under diverse illumination conditions. Our code is publicly available at https://github.com/ShengyuFang/lumird.git.