Unpaired Image Deraining Using Reward-Guided Self-Reinforcement Strategy
| Authors | Yinghao Chen et al. |
| Year | 2026 |
| Field | Computer Vision |
| arXiv | 2605.00719 |
| Download | |
| Categories | cs.CV |
Abstract
Unsupervised deraining has attracted attention for its ability to learn the real-world distribution of rain without paired supervision. However, the lack of strong constraints makes it difficult for the network to converge, especially with the complex diversity of rain degradation. A key motivation is that high-quality deraining results occasionally emerge during training, which can be leveraged to guide the optimization process. To overcome these challenges, we introduce RGSUD (Reward-Guided Self-Reinforcement Unsupervised Image Deraining), comprising two key stages: reward recycling and self-reinforcement (SR) training. For the former stage, we propose an Image Quality Assessment (IQA)-based dynamic reward recycling mechanism that selects optimal derained outputs during training and continuously collects high-quality deraining images. In latter stage, we incorporate these rewards into the model's optimization process, constraining the optimization space and improving alignment between derained outputs and clean images. By leveraging IQA-based self-reinforced loss and dynamically updated rewards, we enhance the quality of synthesized pseudo-paired data and stabilize the optimization. Extensive experiments demonstrate that our method achieves SOTA performance across multiple datasets, including paired synthetic, paired real, and unpaired real images, outperforming existing unsupervised deraining approaches in both subjective and objective IQA metrics. Additionally, we show that the self-reinforcement strategy is adaptable to other unsupervised deraining methods and our deraining framework demonstrates strong generalization across existing supervised deraining networks.
Engineering Breakdown
Plain English
This paper addresses unsupervised image deraining — removing rain streaks from photos without paired clean/rainy image examples. The authors propose RGSUD, which uses an Image Quality Assessment (IQA) mechanism to identify high-quality derained outputs that occasionally appear during training, then recycles these outputs as pseudo-targets to guide further optimization. The key insight is that even without ground truth, the network can learn to recognize good results and reinforce itself toward that quality, solving a major convergence problem in unsupervised deraining where networks struggle with the complexity and diversity of real rain patterns.
Core Technical Contribution
The core novelty is a two-stage reward-guided self-reinforcement framework that operates without paired training data. First, the authors introduce a dynamic reward recycling mechanism that uses image quality assessment to detect and select optimal derained outputs during training — these become pseudo-labels for the next training iteration. Second, they propose a self-reinforcement training strategy that uses these recycled high-quality outputs to provide explicit optimization targets, effectively creating a positive feedback loop where the network bootstraps its own supervision from emergent good solutions. This is fundamentally different from prior unsupervised deraining methods that rely solely on adversarial losses or cycle-consistency without explicit quality-based guidance.
How It Works
The system operates in two main stages. In the reward recycling stage, during each training epoch, the deraining network produces outputs for input rainy images. An IQA module (likely a pre-trained image quality predictor) scores these outputs, identifying which derained images best match criteria for clean images — these high-quality results are cached as pseudo-ground-truth. In the self-reinforcement stage, the network is retrained using both the original unsupervised loss (adversarial or perceptual-based) and an additional supervised loss comparing its outputs to the recycled high-quality pseudo-targets from the previous stage. This creates an iterative refinement loop where the network progressively improves by learning from its own best solutions, rather than being forced to match a single predetermined target. The IQA component acts as the reward signal, continuously identifying which outputs represent convergence toward realistic deraining.
Production Impact
For teams building real-world deraining systems, this approach eliminates the expensive requirement of collecting paired clean/rainy image datasets, which is impractical at scale in production environments. You could deploy RGSUD on raw user-submitted rainy images without manual annotation, directly improving output quality over training iterations. The trade-off is computational cost — the method requires multiple forward passes (one for IQA scoring per image per epoch) and iterative retraining stages, likely increasing training time 2-3x compared to supervised baselines. Integration is relatively straightforward if you already have a deraining network architecture; you mainly need to plug in an IQA module (can be any existing metric like BRISQUE, NIQE, or a learned quality predictor) and restructure your training loop to cache and recycle outputs. This is particularly valuable for mobile or edge deployments where you need to adapt to local rain characteristics without ground truth annotation.
Limitations and When Not to Use This
The paper's reliance on IQA metrics as a reward signal is a fundamental limitation — these metrics may not perfectly correlate with perceptual deraining quality or task-specific requirements (e.g., for autonomous driving, what matters is object visibility, not general sharpness). The method assumes that 'good' derained images will consistently score high on IQA metrics, but adversarial artifacts or texture hallucinations might fool quality assessors while degrading downstream task performance. The iterative recycling process could accumulate errors — if early training produces flawed pseudo-targets, subsequent reinforcement may entrench those mistakes rather than correct them. Additionally, convergence and stability are not guaranteed; without paired supervision, the network could collapse into local minima producing over-smoothed or artifact-heavy outputs that still score well on generic IQA. The paper likely doesn't address how to handle rain with extreme diversity (fog, snow, waterdrops on camera lens) or preserve fine details like text in deraining.
Research Context
This work builds on a decade of unsupervised image deraining research starting with cycle-consistency approaches and adversarial training for rain removal. It directly addresses a known weakness in unsupervised methods — poor convergence without explicit supervision signals — by introducing a self-training mechanism inspired by recent semi-supervised and reinforcement learning trends in vision. The paper likely benchmarks on standard datasets like Rain100L, Rain100H, Test100, and Test2800, comparing against prior unsupervised methods like Unpaired Image Deraining (ID-Cyle), Semi-Supervised Deraining, and recent transformer-based approaches. This opens research directions in quality-guided self-training for other unpaired image restoration tasks (dehazing, denoising, super-resolution) and raises questions about how to design better reward signals beyond generic IQA metrics.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::