Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RetinexDual: Retinex-based Dual Nature Approach for Generalized Ultra-High-Definition Image Restoration

Mohab Kishawy, Ali Abdellatif Hussein, Jun Chen

TL;DR

RetinexDual tackles the challenge of generalized UHD image restoration by decoupling images into reflectance and illumination components and processing them with two specialized branches. The Scale-Adaptive MaMBA (SAMBA) handles reflectance corrections with a coarse-to-fine, nonlocal strategy, while the Frequency Illumination Adaptor (FIA) corrects illumination and color in the Fourier domain using a compact Fourier Correction Block. A multi-level training objective and comprehensive ablations validate the distinct contributions of each branch and their synergy. Across four UHD IR tasks—deraining, deblurring, dehazing, and low-light image enhancement—the method delivers state-of-the-art PSNR/SSIM and superior qualitative results, demonstrating the potential of a general Retinex-based approach for UHD restoration with improved scalability and flexibility.

Abstract

Advancements in image sensing have elevated the importance of Ultra-High-Definition Image Restoration (UHD IR). Traditional methods, such as extreme downsampling or transformation from the spatial to the frequency domain, encounter significant drawbacks: downsampling induces irreversible information loss in UHD images, while our frequency analysis reveals that pure frequency-domain approaches are ineffective for spatially confined image artifacts, primarily due to the loss of degradation locality. To overcome these limitations, we present RetinexDual, a novel Retinex theory-based framework designed for generalized UHD IR tasks. RetinexDual leverages two complementary sub-networks: the Scale-Attentive maMBA (SAMBA) and the Frequency Illumination Adaptor (FIA). SAMBA, responsible for correcting the reflectance component, utilizes a coarse-to-fine mechanism to overcome the causal modeling of mamba, which effectively reduces artifacts and restores intricate details. On the other hand, FIA ensures precise correction of color and illumination distortions by operating in the frequency domain and leveraging the global context provided by it. Evaluating RetinexDual on four UHD IR tasks, namely deraining, deblurring, dehazing, and Low-Light Image Enhancement (LLIE), shows that it outperforms recent methods qualitatively and quantitatively. Ablation studies demonstrate the importance of employing distinct designs for each branch in RetinexDual, as well as the effectiveness of its various components.

RetinexDual: Retinex-based Dual Nature Approach for Generalized Ultra-High-Definition Image Restoration

TL;DR

RetinexDual tackles the challenge of generalized UHD image restoration by decoupling images into reflectance and illumination components and processing them with two specialized branches. The Scale-Adaptive MaMBA (SAMBA) handles reflectance corrections with a coarse-to-fine, nonlocal strategy, while the Frequency Illumination Adaptor (FIA) corrects illumination and color in the Fourier domain using a compact Fourier Correction Block. A multi-level training objective and comprehensive ablations validate the distinct contributions of each branch and their synergy. Across four UHD IR tasks—deraining, deblurring, dehazing, and low-light image enhancement—the method delivers state-of-the-art PSNR/SSIM and superior qualitative results, demonstrating the potential of a general Retinex-based approach for UHD restoration with improved scalability and flexibility.

Abstract

Advancements in image sensing have elevated the importance of Ultra-High-Definition Image Restoration (UHD IR). Traditional methods, such as extreme downsampling or transformation from the spatial to the frequency domain, encounter significant drawbacks: downsampling induces irreversible information loss in UHD images, while our frequency analysis reveals that pure frequency-domain approaches are ineffective for spatially confined image artifacts, primarily due to the loss of degradation locality. To overcome these limitations, we present RetinexDual, a novel Retinex theory-based framework designed for generalized UHD IR tasks. RetinexDual leverages two complementary sub-networks: the Scale-Attentive maMBA (SAMBA) and the Frequency Illumination Adaptor (FIA). SAMBA, responsible for correcting the reflectance component, utilizes a coarse-to-fine mechanism to overcome the causal modeling of mamba, which effectively reduces artifacts and restores intricate details. On the other hand, FIA ensures precise correction of color and illumination distortions by operating in the frequency domain and leveraging the global context provided by it. Evaluating RetinexDual on four UHD IR tasks, namely deraining, deblurring, dehazing, and Low-Light Image Enhancement (LLIE), shows that it outperforms recent methods qualitatively and quantitatively. Ablation studies demonstrate the importance of employing distinct designs for each branch in RetinexDual, as well as the effectiveness of its various components.

Paper Structure

This paper contains 21 sections, 9 equations, 6 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Ultra-High-Definition Image Restoration.
  4. Retinex Theory in Image Restoration.
  5. Methodology
  6. RetinexDual Framework:
  7. Scale Attentive MaMBA (SAMBA):
  8. Scale Adaptive Mamba Block (SAMB).
  9. Group State Space Block (GSSB).
  10. Group State Space Module (GSSM).
  11. Frequency Illumination Adaptor (FIA):
  12. Fourier Correction Block (FCB).
  13. Multi-level Training Objective:
  14. Experiments
  15. Experimental settings:
  16. ...and 6 more sections

Figures (6)

  • Figure 1: Frequency analysis on the difference between distorted and clear images in the dehazing and deblurring task.
  • Figure 1: UHD-LL Quantitative Comparison
  • Figure 2: RetinexDual Overview. Based on retinex theory, it decomposes the UHD image into $R_{eff}$ and $L_{eff}$ and operates on them using 2 sub-networks: Scale Attentive MaMBA (SAMBA), and Frequency Illumination Adaptor (FIA), respectively. Noting that each output from a different level has a convolution layer before it for processing, which wasn't illustrated for simplicity.
  • Figure 3: The architecture of Grouping State Space Block (GSSB)
  • Figure 4: Visual comparison between different architectures on UHD-LL and 4K-Rain 13K. The first row shows results from UHD-LL Li2023ICLR test (1188) while the second row shows results from 4K-Rain13K dataset chen2024towards test (66).
  • ...and 1 more figures