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

2026-03-29

AI Summary

This work addresses the challenging problem of unified restoration for ultra-high-definition (UHD) images degraded by multiple complex factors such as raindrops, low illumination, and noise. The authors propose a physics-guided dual-branch network framework that integrates Retinex decomposition with task-specific physical priors. Central to this approach is a novel Physics-Conditioned Multi-head Self-Attention (PC-MSA) mechanism, which explicitly guides the joint correction of reflectance and illumination components. Notably, this is the first single architecture capable of performing multi-task UHD image restoration without requiring structural modifications across tasks. The model achieved 4th place in the NTIRE 2026 Raindrop Removal Challenge and 5th place in the Joint Noise and Low-Light Enhancement Challenge, demonstrating state-of-the-art performance and computational efficiency.

Abstract

We propose RetinexDualV2, a unified, physically grounded dual-branch framework for diverse Ultra-High-Definition (UHD) image restoration. Unlike generic models, our method employs a Task-Specific Physical Grounding Module (TS-PGM) to extract degradation-aware priors (e.g., rain masks and dark channels). These explicitly guide a Retinex decomposition network via a novel Physical-conditioned Multi-head Self-Attention (PC-MSA) mechanism, enabling robust reflection and illumination correction. This physical conditioning allows a single architecture to handle various complex degradations seamlessly, without task-specific structural modifications. RetinexDualV2 demonstrates exceptional generalizability, securing 4\textsuperscript{th} place in the NTIRE 2026 Day and Night Raindrop Removal Challenge and 5\textsuperscript{th} place in the Joint Noise Low-light Enhancement (JNLLIE) Challenge. Extensive experiments confirm the state-of-the-art performance and efficiency of our physically motivated approach.

Problem

• UHD image restoration
• image degradation
• raindrop removal
• low-light enhancement
• generalized restoration

Innovation

• Physically-Grounded
• Retinex Decomposition
• Dual-Branch Framework
• Degradation-Aware Priors
• Physical-conditioned Multi-head Self-Attention