Scholarlink

<p>Low-light image enhancement aims to improve the visibility and contrast of underexposed images while mitigating distortions such as noise and color artifacts. Traditional methods, often rooted in Retinex theory, tend to suffer from noise amplification and color distortion. Recent advancements in deep learning, especially convolutional neural networks (CNNs), have shown promising results but face challenges in capturing long-range dependencies and non-local self-similarity. To overcome these limitations, we introduce <b>DED-Net</b>, a novel framework that leverages multi-scale features to estimate illumination information and restore image details. Our approach includes an <b>Image Feature Enhancement (IFE) Module</b> that extracts finer feature granularity by sampling the original image at various resolutions, thus guiding the preliminary enhancement. Next, a Retinex-based method decouples the original and enhanced images into illumination features and an illumination map, which are then fused to produce reconstructed and guidance features. Finally, the <b>Illumination-Guided Restoration (IGR) Module</b> uses these illumination representations to model non-local interactions across regions with varying lighting conditions. By integrating depthwise separable convolutions, DED-Net delivers high restoration quality while keeping computational costs low. Experimental evaluations on benchmark datasets (e.g., <b>LOLv1</b> and <b>LOLv2-syn</b>) show that DED-Net outperforms existing methods in both efficiency and restoration quality, achieving a PSNR of 25.87 dB and an SSIM of 0.848 on the LOLv1 dataset. The code for this paper has been released at <a href="https://github.com/Dhs2004/DED-Net.git">https://github.com/Dhs2004/DED-Net.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

DED-net: multi-scale fusion and illumination-guided enhancement for low-light image restoration

  • 期刊论文

  • 发表日期：

    2025年5月30日

查看全文
  • Haosen Dong,
  • Shihao Cheng,
  • Tianyou Fang,
  • Jin Chen,
  • Zhigang Tu

摘要

Low-light image enhancement aims to improve the visibility and contrast of underexposed images while mitigating distortions such as noise and color artifacts. Traditional methods, often rooted in Retinex theory, tend to suffer from noise amplification and color distortion. Recent advancements in deep learning, especially convolutional neural networks (CNNs), have shown promising results but face challenges in capturing long-range dependencies and non-local self-similarity. To overcome these limitations, we introduce DED-Net, a novel framework that leverages multi-scale features to estimate illumination information and restore image details. Our approach includes an Image Feature Enhancement (IFE) Module that extracts finer feature granularity by sampling the original image at various resolutions, thus guiding the preliminary enhancement. Next, a Retinex-based method decouples the original and enhanced images into illumination features and an illumination map, which are then fused to produce reconstructed and guidance features. Finally, the Illumination-Guided Restoration (IGR) Module uses these illumination representations to model non-local interactions across regions with varying lighting conditions. By integrating depthwise separable convolutions, DED-Net delivers high restoration quality while keeping computational costs low. Experimental evaluations on benchmark datasets (e.g., LOLv1 and LOLv2-syn) show that DED-Net outperforms existing methods in both efficiency and restoration quality, achieving a PSNR of 25.87 dB and an SSIM of 0.848 on the LOLv1 dataset. The code for this paper has been released at https://github.com/Dhs2004/DED-Net.git