💻 computer science

What to Remove, What to Preserve: Dual-Ambiguity Rectification for All-in-One Image Restoration

The paper proposes DAR-Net, a novel all-in-one image restoration framework that addresses the challenges of semantic and spatial ambiguity through a structured Degradation Archetype Representation and dedicated rectification modules, achieving state-of-the-art performance across diverse degradation benchmarks.

Original authors: Cencen Liu (University of Electronic Science and Technology of China), Wen Yin (University of Electronic Science and Technology of China), Dongyang Zhang (University of Electronic Science and Technolo

Published 2026-07-31

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Cencen Liu (University of Electronic Science and Technology of China), Wen Yin (University of Electronic Science and Technology of China), Dongyang Zhang (University of Electronic Science and Technology of China), Dongmin Li (University of Electronic Science and Technology of China), Shan Zhao (Jiigan Technology), Bing Su (Jiigan Technology), Tao He (University of Electronic Science and Technology of China), Jielei Wang (University of Electronic Science and Technology of China), Guoming Lu (University of Electronic Science and Technology of China)

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 are a master chef trying to fix a pot of soup that has gone wrong. Maybe it's too salty, maybe it's burnt, or maybe it's full of unwanted weeds. In the world of computer vision, this "soup" is a digital photo, and the "mistakes" are things like rain streaks, fog, blur, or noise. For a long time, scientists built special chefs for each specific problem: one chef just for removing rain, another just for clearing fog. But in the real world, photos often suffer from a messy mix of problems all at once. We need an "All-in-One" chef that can handle any disaster in a single pot. The challenge, however, is that these digital chefs often get confused. When they try to remove the "bad stuff" (like rain), they sometimes accidentally throw away the "good stuff" (like the texture of a person's face or the clouds in the sky). They get tangled up, not knowing what to keep and what to toss. This paper tackles that confusion by teaching the computer exactly how to separate the trash from the treasure.

The paper introduces a new system called DAR-Net, which acts like a super-smart, dual-brained editor for fixing photos. The authors noticed that existing methods suffer from "dual ambiguity," which is a fancy way of saying the computer gets mixed up in two specific ways. First, it gets confused about what the problem is (Semantic Ambiguity), and second, it gets confused about where to fix it (Spatial Ambiguity). To fix this, DAR-Net uses three clever tricks.

First, it uses a Degradation Archetype Representation (DAR). Think of this as a "menu of disasters." Instead of trying to memorize every single possible way a photo can get ruined, the system learns a small set of basic "archetypes" (like a basic rain pattern, a basic fog pattern, etc.). When a new photo comes in, the system figures out which mix of these basic disasters is present, creating a clear "recipe" for what needs to be removed.

Second, it uses Semantic Ambiguity Rectification (SeAR). This is like a smart filter that tells the computer, "Hey, focus on these specific channels of information to remove the rain, but ignore these other channels so you don't delete the clouds." It generates a special "prompt" (a set of instructions) based on the disaster recipe to make sure the computer knows exactly what to target.

Third, and perhaps most importantly, it uses Spatial Ambiguity Rectification (SpAR). This is the "keep vs. remove" referee. Even with the right instructions, the computer might still try to fix the wrong spots. SpAR forces the "remove" signals and the "keep" signals to live in completely separate, non-overlapping spaces. Imagine trying to sort red and blue marbles; SpAR ensures the red ones (the bad stuff) go into one box and the blue ones (the good stuff) go into another, so they never get mixed up again.

The results are impressive. The authors tested DAR-Net on standard benchmarks where photos were ruined by three different types of problems (like rain, fog, and noise) and found it outperformed the best previous methods by an average of 0.14 dB in PSNR (a measure of image quality). When they cranked it up to five different types of problems, it beat the competition by 0.34 dB. It also did a great job on real-world photos with mixed weather issues, showing that it doesn't just work on perfect test data but can handle the messy reality of a rainy, foggy day.

The paper suggests that by explicitly teaching the computer to distinguish between "what to remove" and "what to preserve," rather than letting them get tangled together, we can build much more reliable tools for restoring our digital memories. While the system is complex and requires powerful computers to run, the authors show that this approach of "dual-ambiguity rectification" is a promising path forward for making all-in-one photo fixers that actually work as well as they promise.

Technical Summary: What to Remove, What to Preserve: Dual-Ambiguity Rectification for All-in-One Image Restoration

Problem Statement

All-in-one image restoration (AIR) aims to recover clean visual content from diverse degradations (e.g., noise, haze, rain, blur) within a single unified framework. While existing methods have moved from task-specific models to shared-backbone conditional restoration (often using prompts or Mixture-of-Experts), they face a fundamental limitation: degradation-content entanglement. In current unified pipelines, degradation-related cues and scene content representations are often encoded in an entangled manner within a shared latent space.

The authors characterize this entanglement as dual ambiguity, leading to two typical failure modes:

Semantic Ambiguity (Channel-wise): Degradation cues and content representations are entangled in the channel dimension, causing degradation-aware modulation to be less discriminative. This results in content corruption, where faithful image details are mistakenly suppressed along with the degradation.
Spatial Ambiguity (Spatial dimension): Entanglement in the spatial dimension weakens the selectivity of restoration responses, leading to residual degradation, where artifacts are not sufficiently suppressed.

Methodology: DAR-Net

To address these issues, the authors propose DAR-Net (Dual-Ambiguity Rectification Network), built upon a hierarchical U-shaped Transformer backbone. The network employs a three-stage rectification pipeline to disentangle degradation and content in both channel and spatial dimensions:

Degradation Archetype Representation (DAR):
- Instead of treating degradation as a single vector, DAR constructs a structured degradation state using simplex-constrained archetype mixture modeling.
- A lightweight conditioning network extracts features from the input, which are mapped to mixture coefficients ( $\alpha$ ) via a Softmax operation. These coefficients represent the input as a convex combination of $K$ learnable degradation archetypes ( $A$ ).
- The resulting degradation state ( $S_{deg} = A\alpha$ ) serves as a structured prior for subsequent rectification, ensuring the degradation representation lies within a geometrically constrained space.
Semantic Ambiguity Rectification (SeAR):
- This module addresses channel-wise entanglement using two sub-components:
  - Archetype-Guided Prompt Generator (AGPG): Generates a base prompt from the input features and refines it using the degradation state ( $S_{deg}$ ) via channel-wise routing. This produces a degradation-aware prompt that emphasizes channels consistent with the inferred degradation while suppressing irrelevant ones.
  - Degradation-Aware Prompt Integrator (DAPI): Injects the refined prompt into the decoder stages using channel-wise attention mechanisms to modulate features based on the degradation condition.
- Crucially, at the deepest decoder stage, SeAR derives a complementary content-related feature ( $F_c$ ) by performing residual subtraction ( $F_{in} - F_{dp}$ ), where $F_{dp}$ is the degradation-aware prompt feature.
Spatial Ambiguity Rectification (SpAR):
- To address spatial entanglement, SpAR introduces an Orthogonal Subspace Rectification (OSR) strategy.
- It takes the degradation-aware prompt ( $F_{dp}$ ) and the complementary content feature ( $F_c$ ) and projects them into learnable subspaces.
- A loss term ( $L_{OSR}$ ) is applied to minimize the cross-Gram matrix between the row spaces of these two features. This forces the degradation-related and content-related representations to occupy orthogonal spatial subspaces, effectively separating "what to remove" from "what to preserve."

Training Objective: The network is trained with a pixel-wise L1 reconstruction loss and the orthogonality regularization loss ( $L_{OSR}$ ).

Key Contributions

Dual-Ambiguity Framework: The paper identifies and formalizes the problem of semantic (channel) and spatial ambiguity in AIR, proposing a unified framework to rectify both.
Structured Degradation Modeling: The introduction of the DAR module, which models degradation states as simplex-constrained mixtures of archetypes, providing a more structured prior than standard conditioning vectors.
Orthogonal Separation: The design of the SpAR module, which explicitly enforces orthogonality between degradation and content features in the spatial dimension to prevent interference.
Prompt-Based Rectification: A novel approach to generating degradation-aware prompts that are refined by archetype guidance and integrated via channel attention.

Experimental Results

DAR-Net was evaluated on standard all-in-one restoration benchmarks under three-degradation (3D) and five-degradation (5D) settings, as well as on mixed-degradation (CDD-11) and real-world (WeatherBench) datasets.

3D Setting: DAR-Net achieved the best overall performance with an average PSNR of 33.05 dB and SSIM of 0.922, outperforming the strongest competitor (MIRAGE) by 0.14 dB in average PSNR. It ranked first across dehazing, deraining, and all denoising levels.
5D Setting: In the more challenging 5D setting (including deblurring and low-light), DAR-Net achieved an average PSNR of 31.02 dB and SSIM of 0.921, improving over the second-best method by 0.34 dB.
Mixed & Real-World: On the CDD-11 dataset (covering single to triple degradations), DAR-Net achieved the highest average PSNR/SSIM (29.73/0.888). On the WeatherBench real-world dataset, it demonstrated superior performance in dehazing, desnowing, and deraining, with strong perceptual quality (low LPIPS and FID).
Efficiency: With 35.5M parameters and 771G FLOPs, DAR-Net maintains competitive inference efficiency (165ms latency on 720x480 images), performing faster than several state-of-the-art baselines like PromptIR and DFPIR.

Significance and Claims

The paper claims that the primary contribution of DAR-Net lies in its explicit modeling of "what to remove and what to preserve." By addressing the dual ambiguity that plagues existing unified restoration models, the method effectively mitigates the trade-off between degradation removal and content preservation.

The authors emphasize that their approach does not rely on increasing model capacity through massive parameter counts or complex routing alone, but rather on structural rectification of the feature space. The results suggest that enforcing orthogonality between degradation and content representations is a critical factor in achieving high-fidelity, generalizable all-in-one image restoration. The work provides a new perspective for future unified restoration frameworks, shifting focus from mere feature sharing to the active disentanglement of restoration cues.