💻 computer science

RetinexDualV2: Physically-Grounded Dual Retinex for Generalized UHD Image Restoration

RetinexDualV2 is a unified, physically grounded dual-branch framework that leverages a Task-Specific Physical Grounding Module and a novel Physical-conditioned Multi-head Self-Attention mechanism to achieve state-of-the-art, generalized Ultra-High-Definition image restoration across diverse degradations without requiring task-specific structural modifications.

Original authors: Mohab Kishawy, Jun Chen

Published 2026-03-31

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Mohab Kishawy, Jun Chen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a camera that can take incredibly sharp, high-definition photos (Ultra-High-Definition, or UHD). But sometimes, the world isn't perfect. Your photo might be taken in the dark, covered in rain, foggy, or have weird shadows. Fixing these photos is like trying to restore a damaged painting, but the painting is 4K resolution, and you have to do it fast without breaking your computer.

This paper introduces RetinexDualV2, a new AI tool designed to fix these messy photos. Here is how it works, explained simply:

1. The Core Idea: Separating the "Paint" from the "Light"

To understand the problem, imagine a photo is made of two things:

The Paint (Reflectance): The actual colors and details of the object (like the red of a car or the green of a tree). This should stay the same.
The Light (Illumination): How bright or dark the scene is, or how the weather affects the visibility. This is usually what gets messed up.

Old AI tools often tried to fix the whole photo at once, like trying to clean a muddy window by scrubbing the whole thing. RetinexDualV2 is smarter. It acts like a two-person cleaning crew:

Crew A focuses only on the "Paint" (making sure the colors and details are sharp).
Crew B focuses only on the "Light" (fixing the brightness, fog, or rain).

By splitting the job, they don't get confused and can do a much better, faster job.

2. The Secret Sauce: The "Physical Grounding" Module

Here is the biggest innovation. Most AI models are like students who just memorize answers by looking at thousands of examples. If they see a picture they haven't seen before, they might get stuck.

RetinexDualV2 is different. It comes with a specialized instruction manual for every specific problem. This is called the Task-Specific Physical Grounding Module (TS-PGM).

Think of it like a detective who changes their toolkit based on the crime scene:

If it's raining: The AI puts on "Rain Glasses." It knows exactly what a raindrop looks like and where it usually hides, so it can surgically remove it without smearing the background.
If it's dark: The AI puts on "Night Vision Goggles." It knows that in the dark, noise (grain) usually hides in the shadows, so it cleans the shadows carefully without making them look fake.
If it's foggy: The AI uses "Fog Radar." It knows fog scatters light, so it calculates exactly how to cut through the haze.

Instead of just guessing, the AI uses real-world physics (like how light scatters or how raindrops refract) to guide its cleaning.

3. The Brain: "Physical-Conditioned Attention"

Once the AI knows what the problem is (via the instruction manual above), it needs to know where to look.

Imagine you are editing a video. You have a "Highlight" button. In this AI, the Physical-Conditioned Multi-head Self-Attention (PC-MSA) is that button.

When the AI sees a raindrop, the "Physical Grounding" module shouts, "Hey! Look right here! That's a raindrop!"
The AI's attention mechanism then zooms in specifically on that spot to fix it, while ignoring the rest of the clean image.
This ensures the AI doesn't accidentally blur a sharp tree just because it's trying to remove rain.

4. Why is this a big deal?

One Tool for Many Jobs: Usually, you need one AI for rain, another for fog, and another for darkness. RetinexDualV2 is a Swiss Army Knife. You give it a messy photo, and it automatically figures out which "tool" to use.
It's Efficient: High-definition photos are huge files. Fixing them usually requires a supercomputer. This model is surprisingly lightweight (only about 4.8 million parameters), meaning it can run on standard hardware without melting your laptop.
It Wins Competitions: The authors tested this in the NTIRE 2026 Challenges (a major contest for image restoration). They came in 4th place for rain removal and 5th place for low-light enhancement, beating many much larger and more complex models.

The Bottom Line

RetinexDualV2 is like a master restorer who doesn't just guess how to fix a photo. Instead, they understand the physics of the mess (rain, light, fog), split the job into two specialized teams (fixing light vs. fixing details), and use a smart spotlight to focus only on the damaged areas. The result? Crystal clear, high-definition photos, even when the original conditions were terrible.

1. Problem Statement

The paper addresses the challenge of Ultra-High-Definition (UHD) Image Restoration (IR) under diverse and complex degradation conditions, including low-light, haze, rain, and shadows.

Limitations of Current Methods:
- Computational Cost: Existing high-performance models often require excessive resources, making them impractical for UHD resolutions.
- Information Loss: Downsampling-based approaches discard high-frequency details.
- Domain Specificity: Many methods are tailored to specific degradations (e.g., only low-light or only dehazing) and lack generalizability.
- Theoretical Gaps: While Retinex-based models (e.g., Retinexformer) excel in low-light, they often treat reflectance and illumination as homogeneous or lack explicit physical grounding for broader restoration tasks. Purely data-driven models often fail to produce physically plausible results in extreme conditions.

2. Methodology: RetinexDualV2

The authors propose RetinexDualV2, a unified, physically grounded, dual-branch framework designed to handle multiple restoration tasks within a single architecture without task-specific structural changes.

Core Architecture

The model is based on Retinex Theory, decomposing a degraded image $I$ into effective Reflectance ( $R_{eff}$ ) and Illumination ( $L_{eff}$ ) components, while modeling artifacts ( $\hat{R}, \hat{L}$ ). The restoration process involves:

Dual-Branch Processing:
- Reflectance Branch (PG-SAMBA): Uses a Physical-Grounded Scale Attentive Mamba Block to correct reflectance details. It employs a coarse-to-fine strategy with Group State Space Blocks (GSSB) for long-range consistency.
- Illumination Branch (PG-FIA): Uses a Physical-Grounded Frequency Illumination Adaptor to adjust illumination. It operates in the frequency domain (Fourier transform) to efficiently correct global contrast and color biases with a lightweight parameter footprint (~0.2M).

Key Innovations

Task-Specific Physical Grounding Module (TS-PGM):
- Instead of learning degradation features implicitly, this module explicitly extracts physical priors tailored to the specific degradation type.
- Deraining: Extracts a residual rain intensity mask.
- Low-Light: Extracts a structure-aware illumination prior using Weberized invariant features (robust to lighting changes).
- Dehazing: Computes the Dark Channel Prior to identify haze regions.
- Shadow Removal: Uses a log-chromaticity shadow-resistant representation to isolate shadow cues independent of brightness.
Physical-Conditioned Multi-head Self Attention (PC-MSA):
- This is the core mechanism that integrates the physical priors into the deep learning process.
- Unlike standard attention, PC-MSA modulates the Value vectors in the self-attention mechanism using the extracted physical map ( $P_m$ ).
- Formula: $V' = V \odot P_m$ . This ensures that feature refinement respects the physical constraints of the degradation (e.g., focusing attention on rain streaks or shadow boundaries).

Training Strategy

Multi-level Supervision: The network outputs predictions at three different resolutions.
Loss Function: A combination of Charbonnier loss, SSIM loss, FFT loss (frequency domain), and Perceptual loss (AlexNet features) to ensure both pixel accuracy and structural/texture fidelity.

3. Key Contributions

Unified Framework: A single architecture capable of handling Low-Light Enhancement (LLIE), Dehazing, Deraining, and Shadow Removal simultaneously.
Physical Grounding: Introduction of the TS-PGM which injects domain-specific physical knowledge (e.g., dark channel, rain masks) directly into the network, moving beyond purely statistical learning.
PC-MSA Mechanism: A novel attention mechanism that uses physical maps to modulate feature interactions, ensuring the model learns physically plausible solutions.
Efficiency: The model maintains a compact size of 4.8M parameters, making it efficient for UHD processing compared to heavier competitors (e.g., DRSformer with 33M+ parameters).

4. Experimental Results

The model was evaluated on standard benchmarks and the NTIRE 2026 Challenges.

Benchmark Performance (4K-Rain13k & UHD-LL):
- Deraining: Achieved 34.56 PSNR and 0.973 SSIM, outperforming previous SOTA methods like RetinexDual (34.50 PSNR) and ERR (34.48 PSNR) with a comparable parameter count.
- Low-Light Enhancement: Achieved 28.91 PSNR and 0.935 SSIM on UHD-LL, surpassing RetinexDual (28.86 PSNR) and ERR (27.57 PSNR).
NTIRE 2026 Challenge Rankings:
- Day and Night Raindrop Removal: Secured 4th place (PSNR: 27.24, SSIM: 0.806).
- Joint Noise Low-light Enhancement: Secured 5th place (PSNR: 18.69, SSIM: 0.65).
- Also achieved competitive results in Nighttime Dehazing and Shadow Removal challenges.
Ablation Study:
- Removing the dual-branch structure dropped performance significantly (26.67 PSNR).
- Removing the physical prior reduced performance to 28.79 PSNR.
- Replacing PC-MSA with simple concatenation of priors resulted in 28.43 PSNR.
- Conclusion: The combination of dual-branch decomposition, explicit physical priors, and PC-MSA conditioning is essential for optimal performance.

5. Significance

Generalizability: RetinexDualV2 demonstrates that a single, unified model can effectively handle diverse, complex degradations without needing task-specific architectural re-engineering.
Interpretability: By integrating physical priors, the model moves away from "black-box" deep learning toward interpretable foundation models that adhere to the physics of image formation.
Efficiency: It achieves state-of-the-art results with a relatively small parameter count, addressing the critical bottleneck of computational cost in UHD image restoration.
Robustness: The method proves highly resilient in real-world scenarios (e.g., dual-focused raindrop removal, joint noise/low-light), suggesting strong potential for deployment in autonomous driving, surveillance, and consumer photography.