Research map / Image generation / Image restoration
Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CSUD, an unsupervised image deraining framework that trains on unpaired data by generating pseudo rainy-clean pairs, using a channel consistency loss grounded in a channel consistency prior of rain streaks and a self-reconstruction strategy.
Problem
Image deraining without paired real rainy-clean data, where existing unsupervised methods lag supervised ones and both generalize poorly to real-world rain.
Approach
Introduces a channel consistency prior (cycle subtractions across R,G,B are rain-free and consistent between clean and rainy images) as a CCLoss for the generator, plus a self-reconstruction loss that makes generator and derainer reconstruct the clean input from itself, reducing redundant background transfer into pseudo rainy images.
Key results
- CSUD reaches 33.28 dB PSNR on Rain100L and 32.71 dB on RealRain1K-L, outperforming the best unsupervised baseline DCD-GAN by 1.46 dB on Rain100L and 2.22 dB on RealRain1K-L.
- In generalization trained on Rain100L, CSUD scores 29.21 dB on RealRain1K-L and 33.57 dB on SPA-data, surpassing supervised NAFNet, PromptIR and Restormer on the first four real-world datasets.
- Ablation shows CCLoss and SR strategy together raise Rain100L PSNR from 30.66 to 33.28 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CSUD instead of DCD-GAN | Unsupervised deraining · Rain100L | PSNR: 31.82 -> 33.28; 1.46dB improvement | Helped |
| Use CSUD instead of DCD-GAN | Unsupervised deraining · Rain100H | PSNR: 22.47 -> 24.42 | Helped |
| Use CSUD instead of DCD-GAN | Unsupervised deraining · Rain12 | PSNR: 31.56 -> 34.56; 3.00dB improvement | Helped |
| Use CSUD instead of DCD-GAN | Unsupervised deraining · Rain800 | PSNR: 25.61 -> 23.58 | Hurt |
| Use CSUD instead of DCD-GAN | Unsupervised deraining · RealRain1K-L | PSNR: 30.49 -> 32.71; 2.22dB improvement | Helped |
| Add CCLoss to the baseline without SR strategy (ablation) | CSUD · Rain100L | PSNR: 30.66 -> 31.24; 1.38dB | Helped |
| Add CCLoss to the baseline without SR strategy (ablation) | CSUD · Rain12 | PSNR: 31.60 -> 32.62; 1.02dB | Helped |
| Add CCLoss to the baseline without SR strategy (ablation) | CSUD · RealRain1K-L | PSNR: 30.42 -> 31.14; 0.72dB | Helped |
| Add SR strategy to the baseline without CCLoss (ablation) | CSUD · Rain100L | PSNR: 30.66 -> 32.69; 1.38dB | Helped |
| Add SR strategy to the baseline without CCLoss (ablation) | CSUD · Rain12 | PSNR: 31.60 -> 33.43; 1.83dB | Helped |
| Add SR strategy to the baseline without CCLoss (ablation) | CSUD · RealRain1K-L | PSNR: 30.42 -> 31.86; 1.44dB | Helped |
| Add both CCLoss and SR strategy to the baseline (ablation) | CSUD · Rain100L | PSNR: 30.66 -> 33.28; 2.62dB | Helped |
Surprising findings
- Supervised models trained on synthetic data such as Restormer and NeRD-Rain-S perform poorly on real-world rain, while the unsupervised CSUD generalizes better on several real datasets.
Limitations
- Performance on the nighttime dataset is only competitive with, not better than, supervised baselines.
- Underperforms on Rain100H and Rain800 relative to some supervised methods; generalization from Rain100H is weakest.
Experimental setup
Datasets: Rain100L Rain100H Rain12 Rain800 RealRain1K-L RealRain1K-H SPA-data RainDS Internet-Data Night-Rain
Benchmarks: Rain100L Rain100H Rain12 Rain800 RealRain1K-L RealRain1K-H SPA-data RainDS Night-Rain
Metrics: PSNR SSIM
Baselines: CycleGAN DerainCycleGAN NLCL DCD-GAN DerainNet DDN SPA-Net MPRNet NAFNet Restormer PromptIR NeRD-Rain-S
Builds on: CycleGAN DerainCycleGAN NAFNet PatchGAN
Closest papers by meaning
- Unsupervised Deraining: Where Contrastive Learning Meets Self-Similarity (CVPR 2022 · similarity 0.90)
- Unpaired Image Deraining Using Reward-Guided Self-Reinforcement Strategy (CVPR 2026 · similarity 0.88)
- Multi-Decoding Deraining Network and Quasi-Sparsity Based Training (CVPR 2021 · similarity 0.88)
- UniRain: Unified Image Deraining with RAG-based Dataset Distillation and Multi-objective Reweighted Optimization (CVPR 2026 · similarity 0.88)
- Unpaired Deep Image Deraining Using Dual Contrastive Learning (CVPR 2022 · similarity 0.87)
- Rethinking Multi-Scale Representations in Deep Deraining Transformer (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.