Scholarlink

<p>Low-light conditions often degrade image quality, posing challenges in various multimedia and computer vision applications. Based on the Retinex theory, we introduce components-balanced guided diffusion (CBGD), a unibranch Retinex-diffusion framework that regards the normal-light condition as the balanced state of reflectance-illumination components. The proposed CBGD concatenates reflectance and illumination into a multi-channel input and performs a single diffusion process, enforcing shared, synergistic corrections across both components at every timestep. To direct the reverse process toward the normal-light state, we design a sampling-based relighting guidance to provide low-light-to-normal-light supervision. Furthermore, we introduce multiple losses that constrain reflectance, illumination, and their recomposed image jointly. Extensive experiments on 10 public datasets demonstrate state-of-the-art performance, with significant improvements in LPIPS, PSNR, SSIM, and NIQE metrics. When applied as preprocessing for low-light object detection, CBGD significantly boosts detection accuracy, confirming its practical utility. The code and results are available at: <a href="https://github.com/Klawens/CBGD">https://github.com/Klawens/CBGD</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

Components-balanced guided diffusion for retinex-based low-light image enhancement

  • 期刊论文

  • 发表日期：

    2026年2月13日

查看全文
  • Shichang Liu,
  • Xu Xu,
  • Yanli Liu,
  • Guanyu Xing

摘要

Low-light conditions often degrade image quality, posing challenges in various multimedia and computer vision applications. Based on the Retinex theory, we introduce components-balanced guided diffusion (CBGD), a unibranch Retinex-diffusion framework that regards the normal-light condition as the balanced state of reflectance-illumination components. The proposed CBGD concatenates reflectance and illumination into a multi-channel input and performs a single diffusion process, enforcing shared, synergistic corrections across both components at every timestep. To direct the reverse process toward the normal-light state, we design a sampling-based relighting guidance to provide low-light-to-normal-light supervision. Furthermore, we introduce multiple losses that constrain reflectance, illumination, and their recomposed image jointly. Extensive experiments on 10 public datasets demonstrate state-of-the-art performance, with significant improvements in LPIPS, PSNR, SSIM, and NIQE metrics. When applied as preprocessing for low-light object detection, CBGD significantly boosts detection accuracy, confirming its practical utility. The code and results are available at: https://github.com/Klawens/CBGD.