Degradation-Agnostic Clarity Learning for Unpaired Image Dehazing¶
Conference: ECCV 2026
Paper: ECCV Official
Area: Image Restoration
Keywords: Unpaired Image Dehazing, Degradation-Agnostic Clarity Learning, Multi-Task Learning, Pareto Optimization, Self-Supervised Texture Enhancement
TL;DR¶
Addressing severe semantic hallucinations caused by the cross-domain semantic gap in unpaired image dehazing, this paper proposes Degradation-Agnostic Clarity Learning (DCL), which formulates clarity extraction as multi-task Pareto optimization over adversarially mined degradations and integrates an Enhance-to-Resist self-supervised objective for detail recovery.
Background & Motivation¶
In real-world image dehazing, acquiring large-scale paired datasets of hazy and clean images in identical real environments is notoriously difficult. Traditional physical-prior methods rely on simplified atmospheric scattering models, which often fail under complex illumination and weather conditions, leading to noticeable color casts, incomplete haze removal, and severe visual artifacts. Recent advances approach dehazing from an unsupervised domain translation perspective (e.g., using diffusion-based models like CycleGAN-Turbo), bypassing brittle physical assumptions and demonstrating superior adaptability to diverse natural scattering.
However, unsupervised domain translation frameworks frequently induce severe semantic hallucinations and structural distortions, especially in heavily degraded regions where scene contours are obscured. Because unpaired datasets exhibit substantial semantic distribution mismatches between the hazy and clean domains, standard discriminators quickly overfit to high-level semantic features rather than low-level clarity cues. Driven by the global domain-alignment objective, generators hallucinate artificial semantic content to satisfy the discriminator, compromising original scene geometry. Manually rebalancing dataset semantics is labor-intensive and unscalable, making it essential to guide the discriminator strictly toward intrinsic low-level clarity.
Steering unsupervised networks toward pure clarity presents two primary hurdles: suppressing strong interference from high-level semantics, and defining a robust clarity objective without ground-truth reference pairs. Core idea: explicitly inject diverse low-level degradations to establish surface statistical shortcuts that naturally suppress high-level semantic exploitation, and formulate clarity learning as multi-task Pareto optimization across diverse degraded variants to isolate a common clarity gradient direction that eliminates degradation-specific bias and semantic hallucinations.
Method¶
Overall Architecture¶
DCL is built upon an unpaired image-to-image translation backbone comprising a forward dehazing generator \(G_{X \to Y}\), a reverse haze synthesizer \(G_{Y \to X}\), and two domain discriminators \(D_Y\) and \(D_X\). At inference time, only the generator \(G_{X \to Y}\) is executed. The training framework consists of two main modules: Degradation-Shared Clarity Feature Learning (DCFL), which reframes discriminator training as multi-task optimization using adversarial reinforcement learning to mine challenging degradation variants and Pareto optimization (MGDA) to extract the shared clarity trajectory; and Enhance-to-Resist (E2R) Self-Supervised Texture Enhancement, which introduces stochastic high-frequency perturbations into the cycle-consistency loop to enforce texture reconstructability in latent space.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Clean / Hazy Image Input"] --> B["Adversarial AutoAugment Degradation Mining<br/>Policy network explores hard degradation variants"]
B --> C["Pareto Optimization Gradient Decoupling<br/>MGDA extracts shared clarity descent trajectory"]
C --> D["Enhance-to-Resist Self-Supervised Loop<br/>Latent texture consistency under perturbation"]
D --> E["High-Fidelity Dehazed Output<br/>Suppressed hallucinations and preserved textures"]
Key Designs¶
1. Adversarial AutoAugment Degradation Mining: dynamically exposing under-fitted degradation distributions
Using a static set of low-level degradations quickly leads to discriminator saturation, diminishing the capacity of surface statistics to suppress high-level semantic cues. To provide persistent, informative low-level shortcuts, the authors introduce a policy network \(\pi(\tau|\theta_\pi)\) inspired by Adversarial AutoAugment. The policy dynamically samples degradation operations (including noise, rain, blur, sharpen, grayscale, darkening, and JPEG compression) along with their corresponding probabilities and magnitudes from a predefined space \(\mathcal{T}\). The policy network is optimized via reinforcement learning to maximize the discriminator's loss on degraded clean samples: $\(R(\tau) = \mathcal{L}_{D_Y}(T_\tau(y_c); \theta_{D_Y})\)$ By continuously mining \(K\) hard degradation variants that maximally challenge the discriminator's current decision boundary, the system ensures rich degradation diversity and prevents overfitting to a fixed collection of artifacts.
2. Pareto Optimization Clarity Extraction: extracting intrinsic clarity manifolds from conflicting degradations
Given \(K\) mined degradation tasks, simple scalar loss summation allows tasks with dominant gradient magnitudes to dictate optimization, biasing the discriminator toward specific artifact patterns. Grounded in the hypothesis that distinct degraded variants share a common underlying clarity vector, discriminator parameter updates are cast as multi-objective optimization. The Multiple Gradient Descent Algorithm (MGDA) is employed to find the minimum-norm point within the convex hull of task gradients. Let \(\mathcal{L}_k(\theta_{D_Y})\) be the loss for the \(k\)-th degradation task; the optimal task weighting vector \(\boldsymbol{\alpha}\) is solved over the simplex \(\Delta^{K-1}\): $\(\min_{\boldsymbol{\alpha} \in \Delta^{K-1}} \left\| \sum_{k=1}^K \alpha_k \nabla_{\theta_{D_Y}} \mathcal{L}_k(\theta_{D_Y}) \right\|_2^2\)$ Under the zero-mean assumption for diverse degradation vectors, task-specific gradient components cancel out, leaving the final update \(\Delta \theta_{D_Y} = \sum_{k=1}^K \alpha_k \nabla \mathcal{L}_k\) dominated entirely by the shared clarity component, steering the generator away from high-level semantic shortcuts.
3. Enhance-to-Resist Self-Supervised Texture Enhancement: structure preservation under high-frequency perturbations
Unpaired dehazing often lacks pixel-level supervision, causing fine textural details to be smoothed out. Extending the relative clarity concept, the Enhance-to-Resist (E2R) scheme posits that inherent scene textures must remain reconstructable even after undergoing unknown high-frequency corruption. Utilizing a pre-trained, frozen Variational Autoencoder (VAE) encoder \(E\) as a latent structural descriptor, a stochastic high-frequency degradation operator \(T_{hf}\) is applied to the generator output \(y_{\text{gen}} = G_{X \to Y}(x)\) before backward synthesis \(G_{Y \to X}\). The cycle-consistency constraint is evaluated in the VAE latent space: $\(\mathcal{L}_{\text{E2Rcycle}} = \mathbb{E}_{x \sim p_{\text{haze}}, T \sim p(\mathcal{T})} \left[ \| E(x) - E(x_{\text{rec}}^T) \|_2^2 \right]\)$ Minimizing this objective compels the generator to embed robust, input-grounded high-frequency textures capable of surviving structural corruptions, achieving self-supervised texture enhancement without clean reference images.
Loss & Training¶
The overall learning objective integrates unsupervised domain translation constraints and the proposed regularizations: $\(\mathcal{L}_{\text{total}} = \lambda_{\text{GAN}} \mathcal{L}_{\text{GAN}} + \lambda_{\text{cycle}} \mathcal{L}_{\text{cycle}}^{L1} + \lambda_{\text{E2R}} \mathcal{L}_{\text{E2Rcycle}} + \lambda_{\text{idt}} \mathcal{L}_{\text{idt}}\)$ where the adversarial loss on \(D_Y\) aggregates gradients weighted by Pareto optimization. The hyper-parameters are configured as \(\lambda_{\text{cycle}}^{L1} = 1\), \(\lambda_{\text{E2R}} = 10\), \(\lambda_{\text{idt}} = 1\), and \(\lambda_{\text{GAN}} = 0.5\). The degradation candidate pool size is set to \(K = 3\). The model is optimized using AdamW with a learning rate of \(1\times 10^{-5}\) and batch size of 1 on a single NVIDIA RTX 4090 GPU for 30k training iterations.
Key Experimental Results¶
Main Results¶
Quantitative evaluations are performed on real-world dehazing benchmarks including RTTS, Fattal's dataset, and FoggyDriving against leading physical-prior models (Dehamer, KANet, D4+) and modern unpaired translation methods (CycleGAN-Turbo, Diff-Dehazer, DehazeSB). Metrics include no-reference haze density HazDesNet (lower is better) and perceptual quality metrics CLIPIQA, MUSIQ, and MANIQA (higher is better).
| Dataset | Method | HazDesNet↓ | CLIPIQA↑ | MUSIQ↑ | MANIQA↑ |
|---|---|---|---|---|---|
| RTTS | Dehamer | 0.319 | 0.347 | 50.55 | 0.279 |
| RTTS | KANet | 0.372 | 0.280 | 54.53 | 0.257 |
| RTTS | D4+ | 0.333 | 0.305 | 53.29 | 0.289 |
| RTTS | CycleGAN-Turbo | 0.372 | 0.289 | 56.63 | 0.286 |
| RTTS | Diff-Dehazer | 0.321 | 0.356 | 57.80 | 0.271 |
| RTTS | DehazeSB | 0.438 | 0.377 | 53.09 | 0.263 |
| RTTS | Ours | 0.278 | 0.325 | 57.47 | 0.278 |
| Fattal | CycleGAN-Turbo | 0.163 | 0.530 | 66.65 | 0.360 |
| Fattal | Diff-Dehazer | 0.154 | 0.467 | 61.66 | 0.311 |
| Fattal | Ours | 0.098 | 0.587 | 68.83 | 0.438 |
| FoggyDriving (Web) | CycleGAN-Turbo | 0.387 | 0.367 | 58.00 | 0.307 |
| FoggyDriving (Web) | Ours | 0.240 | 0.460 | 60.39 | 0.312 |
Ablation Study¶
Stepwise ablation on the RTTS benchmark systematically isolates the contributions of Adversarial AutoAugment (ARL), Pareto Optimization (PO), and Enhance-to-Resist (E2R):
| Config | ARL | PO | E2R | HazDesNet↓ | CLIPIQA↑ | MUSIQ↑ | MANIQA↑ |
|---|---|---|---|---|---|---|---|
| Baseline (CycleGAN-Turbo) | ✗ | ✗ | ✗ | 0.372 | 0.289 | 56.63 | 0.286 |
| + ARL Degradation Mining | ✓ | ✗ | ✗ | 0.282 | 0.292 | 53.38 | 0.272 |
| ARL + Pareto Optimization (PO) | ✓ | ✓ | ✗ | 0.299 | 0.264 | 54.67 | 0.243 |
| Full Model (DCL) | ✓ | ✓ | ✓ | 0.278 | 0.325 | 57.47 | 0.278 |
Analysis of degradation pool size \(K\): setting \(K = 1\) degenerates to single-task descent, risking degradation-specific overfitting; setting \(K = 3\) provides balanced cross-degradation constraints that cancel out spurious cues; increasing to \(K = 5\) introduces redundant easy tasks that perturb the optimal descent direction, validating \(K = 3\) as the optimal trade-off.
Key Findings¶
- Injecting low-level degradations substantially drops HazDesNet from 0.372 to 0.282, confirming that low-level statistical shortcuts effectively accelerate and reinforce haze-clearing capabilities. Applying Pareto optimization suppresses degradation-specific variance and ensures artifact-free clarity recovery.
- In semantic segmentation downstream evaluations on FoggyDriving using DeepLabV3, DCL achieves 85.5% pxAcc, 50.9% clsAcc, and 37.7% mIoU, clearly outperforming CycleGAN-Turbo (34.9% mIoU) and Diff-Dehazer (33.3% mIoU). This confirms that DCL reliably preserves true scene geometry while mitigating hallucinated artifacts.
- Generalization testing demonstrates that the E2R self-supervised objective transfers seamlessly to low-light image enhancement, adaptively restoring fine textural boundaries across varying illumination levels without amplifying noise.
Highlights & Insights¶
- Reverses the role of adversarial perturbations to construct deliberate surface statistical shortcuts, exploiting the inductive bias of neural networks toward low-level statistics to suppress high-level semantic hallucination.
- Elegantly formulates unsupervised clarity feature learning as a multi-objective Pareto front exploration, leveraging theoretical gradient orthogonality to neutralize task-specific biases.
- Proposes Enhance-to-Resist, moving beyond naive pixel-wise cycle consistency by testing latent structural resilience against adversarial degradation, offering a versatile paradigm for unpaired low-level vision.
Limitations & Future Work¶
- Under severe real-world sensor noise or harsh compression artifacts, the high-frequency reconstruction requirement in E2R may inadvertently treat existing noise patterns as valid structures, slightly over-enhancing residual noise.
- Computing Pareto gradients requires \(K\) separate backward passes through the discriminator per iteration, increasing discriminator training time approximately by a factor of \(K\).
- Future extensions could incorporate noise-level gating mechanisms and expand degradation-agnostic clarity modeling to underwater image restoration, desnowing, and unified all-in-one restoration systems.
Related Work & Insights¶
- vs CycleGAN-Turbo [39]: CycleGAN-Turbo relies on one-step diffusion generative priors for domain mapping but lacks explicit clarity guidance, leading to frequent semantic hallucinations; DCL injects multi-task degradation shortcuts and Pareto balancing to anchor the discriminator strictly on low-level restoration.
- vs Diff-Dehazer [28] & DehazeSB [29]: Existing methods blend physical priors with diffusion models or Schrödinger bridges, but their brittle physical assumptions cause color shifts when atmospheric models fail in real scenes; DCL operates entirely free from rigid physical constraints.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Formulates unpaired clarity learning via multi-task Pareto optimization and statistical shortcuts, offering a fresh, mathematically grounded perspective.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Thorough validations spanning real-world benchmarks, ablations, gradient projection analyses, downstream segmentation, and low-light generalization.
- Writing Quality: ⭐⭐⭐⭐⭐ Well-structured narrative with rigorous mathematical derivations and coherent empirical validations.
- Value: ⭐⭐⭐⭐⭐ Provides a foundational, highly transferable framework for unsupervised image restoration and low-level representation learning.