Real-World Adverse Weather Image Restoration via Dual-Level Reinforcement Learning with High-Quality Cold Start

AI summary

Overview

• Research area: Computer vision, specifically real-world image restoration under adverse weather (rain, haze, snow, and mixtures), combining physics-based synthetic data generation with reinforcement learning and multi-agent coordination.
• Technical level: Advanced. The paper assumes familiarity with image restoration architectures, reinforcement learning (GRPO, PPO-style policy gradients), no-reference image quality assessment metrics (LIQE, CLIP-IQA, Q-Align, MUSIQ), and multi-agent scheduling.
• Scope (one sentence): The paper builds a one-million-image physics-driven weather dataset (HFLS-Weather) for supervised "cold start" training, then refines weather-specific restoration models via a perturbation-based reinforcement learning method and coordinates them with a meta-controller, reporting state-of-the-art real-world results across snow, haze, and rain.

What This Paper Is About

Models trained on synthetic weather data often fail on real photographs because synthetic degradations do not match the physics and variability of actual weather, and because a single fixed model cannot adapt to new degradation patterns. The paper's goal is a self-evolving system that is first "cold started" on a high-fidelity synthetic dataset and then continuously refined on unlabeled real-world images using reinforcement learning, with a higher-level controller that decides which specialized restoration model to run, in what order, for any given scene.

Key Contributions

1. HFLS-Weather dataset: A physics-driven, high-fidelity dataset built from one million clean images collected from Snow100K, RESIDE-OTS, Google Landmark V2, and OSV5M. Depth maps are produced with DepthAnything v2 rather than LiDAR, and a unified atmospheric scattering formulation applies depth-dependent attenuation to haze, rain, and snow, including mixed conditions (rain+haze, snow+haze). The paper reports one million clean images and one million generated pairs.
2. PIQO (Perturbation-driven Image Quality Optimization): A GRPO-inspired local-level reinforcement learning method for image restoration that injects Gaussian perturbations into model parameters to create multiple outputs for a single degraded input, scores them with a composite no-reference reward, filters low-reward outputs using MUSIQ, and updates parameters with normalized advantages plus implicit KL regularization.
3. Global-level multi-agent system: A CLIP-based meta-controller generates a "weather description," broadcasts it to specialized restoration agents that bid based on historical success rates, selects the top-ranked agent, and validates each restoration using CLIP re-analysis and an IQA score, reverting failures and trying the next agent.
4. First GRPO-style application to image restoration: The authors state that this is the first work to successfully apply GRPO concepts to image restoration, and attribute success to the high-quality cold start plus the reward design.

Main Findings

• Highest IQA scores across all three weather types: On the WReal-based evaluation, the method reports Snow Q-Align 3.9569, CLIP-IQA 0.5918, LIQE 3.9458, MUSIQ 67.7990; Haze Q-Align 3.5608, CLIP-IQA 0.4561, LIQE 3.0267, MUSIQ 63.3000; Rain Q-Align 4.0283, CLIP-IQA 0.5623, LIQE 3.2945, MUSIQ 64.1187, outperforming Chen et al., WGWS, PromptIR, OneRestore, and DA-CLIP.
• Best GPT-4o perceptual scores: Snow artifact removal 4.421, weather resilience 4.355, overall visual quality 4.393; Haze 4.074 / 4.015 / 3.948; Rain 4.254 / 4.007 / 3.896, compared against Chen et al., WGWS, PromptIR, OneRestore, DA-CLIP, DFPIR, and JarvisIR.
• Stable superiority across conditions: Competing methods fluctuate by weather type, while the proposed approach maintains consistently higher scores, which the authors attribute to real-world refinement via dual-level reinforcement learning.
• Latency trade-off: The multi-agent system runs at 570 ms, slower than single-model baselines (OneRestore 17 ms, Chen et al. 18 ms, WGWS 95 ms, PromptIR 208 ms) but far faster than other multi-agent systems (DA-CLIP 6543 ms, JarvisIR 15250 ms).
• High-quality cold start matters: Cold-start ablation shows Q-Align of 3.8482 (Our Snow) and 3.8693 (Our Snow+Haze) versus 3.7490 (Snow100K) and 3.6974 (RealSnow); for haze, 3.5329 (Our Haze) versus 3.1220 (OTS) and 3.1014 (ITS); for rain, 3.9318 (Our Rain) and 3.9205 (Our Rain+Haze) versus 3.7805 (SPA+) and 3.6974 (Rain1300).
• Cross-weather pretraining helps: Mixing degradation types during pretraining (e.g., snow+haze) improves performance, suggesting multi-degradation exposure aids generalization during PIQO finetuning.
• Both components are complementary: Ablation shows Basic + PIQO + Agent gives the best scores across snow, haze, and rain (e.g., snow CLIP-IQA 0.5918, Q-Align 3.9458, LIQE 3.9569), better than Basic alone, Basic + Agent, or Basic + PIQO.

Methodology in Plain English

The pipeline has three stages. First, the authors generate training data: they take one million clean photos, predict a depth map for each with DepthAnything v2, and composite synthetic rain, haze, or snow using an atmospheric scattering equation where fog transmission is F(x) = e^{-β d(x)} and rain/snow is a semi-transparent procedural layer M(x). Distant objects are more fogged; near objects show stronger rain streaks or scattered snowflakes.

Second, they train weather-specific restoration models (derain, dehaze, desnow) on this dataset as a supervised "cold start," using a DSANet-based multi-agent system on eight NVIDIA RTX 4090 GPUs, the Adam optimizer, batch size 8, learning rate 0.0001, up to 100 epochs with early stopping on validation loss, plus fine-tuning of the rain sub-model on SPA+.

Third, since real weather photos lack paired clean targets, they refine each model with PIQO. Rather than generating one output per input (as image restoration normally does), they add small Gaussian perturbations to the network weights so that one degraded image yields several different restorations. Each restoration is scored by a weighted no-reference reward combining LIQE, CLIP-IQA, and Q-Align, with weights w1 = 0.2, w2 = 1, and w3 = 0.2. Outputs scoring below the unperturbed model under MUSIQ are discarded, the survivors get normalized advantage scores relative to the group mean, and the weights are nudged toward high-reward perturbations. An implicit KL term, with a threshold τ and a scaling factor, keeps updates from moving too far in one step.

On top of this, the global meta-controller uses CLIP to describe the degradation, collects bids from agents based on their past success with similar situations, and runs the highest-ranked agent. If the resulting IQA score drops, the change is reverted, the agent is removed from the candidate list, and the next agent tries. The system stops when no degradation is detected, after three consecutive failures, or after three agents have been used, returning the highest-IQA image it has seen.

Training uses real data: 2,318 hazy images from the URHI dataset, 2,433 rainy images and 2,018 snowy images from WReal, with learning rate 0.0001 and batch size 16. The global multi-agent stage also uses batch size 16 and learning rate 0.0001. Evaluation uses WReal, which contains 4,322 real haze images from RTTS, 2,320 real rain images from DDN-SIRR and Real3000, and 1,329 real snow images from Snow100K.

Why This Matters

The paper targets a widely reported gap: restoration models that look good on synthetic benchmarks but fail on real photographs. Its main impact is showing that a carefully generated dataset can supply the "cold start" that reinforcement learning needs, and that unlabeled real images can then carry training forward without paired ground truth — a pattern borrowed from language-model alignment and here adapted to image restoration.

Real-world applications (as motivated or implied by the paper):

• Autonomous driving and outdoor robotics in rain, snow, and fog, where degraded visual perception directly affects downstream decisions.
• Surveillance and monitoring systems operating in adverse weather (the authors also flag misuse risk here).
• Computational photography, restoring photos taken in bad weather without a clean reference.
• Safety-critical visual pipelines where geometric and photometric fidelity matters and generative editing tools like GPT-4o may produce hallucinated objects or distorted structure.

Industry relevance: The work comes from a collaboration spanning Nanjing University of Science and Technology, Huawei Noah's Ark Lab, and South China University of Technology, and the method uses a practical multi-agent design with defined latency behavior (570 ms), which matters for deployment trade-offs. The authors also raise societal risks, including misuse for surveillance or deepfake generation and the need for safeguards in safety-critical use.

Future Directions

• Reducing inference overhead: The conclusion explicitly notes that the multi-agent system adds extra inference-time overhead from multi-round interactions; lowering the 570 ms cost is an open engineering question.
• Extending PIQO to more degradation types and combinations: The dataset covers rain, haze, snow, and mixed cases, but whether the perturbation-reward loop scales to other real-world corruptions (motion blur, low light, compression) is untested here.
• Stronger reward design: The reward relies on no-reference IQA models; the paper notes such metrics can fail to capture subtle noise and residual artifacts and can inflate scores, so better reward signals could directly improve PIQO.
• More thorough real-world benchmarking: The evaluation uses unlabeled real datasets judged by IQA and GPT-4o; the paper does not report paired real ground-truth accuracy, so an explicit note is that such results are not reported.

Target Audience

• Researchers and graduate students in image restoration, adverse weather removal, and low-level vision.
• Reinforcement learning researchers interested in how GRPO-style methods transfer from language to continuous image outputs without paired labels.
• Multi-agent and tool-orchestration researchers, since the meta-controller/specialist-agent bidding design is a distinct scheduling approach.
• Industry engineers building perception or camera pipelines that must operate outdoors in bad weather, who need to weigh the reported quality gains against the 570 ms latency.
• Dataset builders, because HFLS-Weather's depth-driven, multi-weather synthesis pipeline is described in enough detail to be reused or compared against.

Authors’ abstract

Adverse weather severely impairs real-world visual perception, while existing vision models trained on synthetic data with fixed parameters struggle to generalize to complex degradations. To address this, we first construct HFLS-Weather, a physics-driven, high-fidelity dataset that simulates diverse weather phenomena, and then design a dual-level reinforcement learning framework initialized with HFLS-Weather for cold-start training. Within this framework, at the local level, weather-specific restoration models are refined through perturbation-driven image quality optimization, enabling reward-based learning without paired supervision; at the global level, a meta-controller dynamically orchestrates model selection and execution order according to scene degradation. This framework enables continuous adaptation to real-world conditions and achieves state-of-the-art performance across a wide range of adverse weather scenarios. Code is available at https://github.com/xxclfy/AgentRL-Real-Weather