Research map / Image generation / Image restoration
Towards Universal Rainy Image Restoration: Benchmark and Baseline
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.
Introduces Universal Rainy Image Restoration (URIR), the URIR-8K benchmark covering four rain patterns with object detection labels, and a multi-scale Mamba baseline (MSDM) with multi-scale 2D scanning.
Problem
Existing deraining methods handle only one rain type or pattern, while real driving scenes mix rain streaks, raindrops, rain accumulation and nighttime rain; no unified benchmark or universal model exists for this.
Approach
Synthesize a unified four-pattern rain dataset on a common driving background set (BDD100K) with object detection labels, and build a multi-scale Mamba network whose 2D scanning uses different numbers of scan directions per image scale.
Key results
- On URIR-8K, the proposed MSDM reaches 30.8 average... specifically 39.35 dB average PSNR and 0.9844 SSIM, exceeding the second-best IDT (37.96 dB) by 1.4 dB.
- MSDM achieves 0.402 mAP50 for downstream object detection on URIR-8K, the best in Table 2 (IDT 0.401).
- MSDM uses 11.8M parameters and 99.2G FLOPs at 256x256, versus 26.1M/174.7G for Restormer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MSDM instead of IDT | URIR-8K universal deraining · URIR-8K average | PSNR: 37.96 -> 39.35 | Helped |
| use MSDM instead of DRSformer | rain accumulation restoration · URIR-8K RA | PSNR: 26.32 -> 29.04 | Hurt |
| add multi-scale 2D scanning instead of fixed-scale 2D scanning (ablation) | MSDM · URIR-8K | PSNR: 40.58 -> 40.89 | Helped |
| add FGCM (ablation) | MSDM · URIR-8K | PSNR: 40.68 -> 41.37 | Helped |
| add MGFM (ablation) | MSDM · URIR-8K | PSNR: 41.16 -> 41.37 | Helped |
| add frequency reconstruction loss Lf and edge loss Le (ablation) | MSDM training · URIR-8K | PSNR: 41.09 -> 41.37 | Helped |
| add edge loss Le (ablation) | MSDM training · URIR-8K | PSNR: 41.27 -> 41.37 | Helped |
| use four-direction scanning at every scale instead of scale-dependent scanning (ablation) | MS-2DSB · URIR-8K | PSNR: 40.89 -> 40.59 | Hurt |
| use all four scanning directions at every scale instead of scale-dependent scanning (ablation) | MS-2DSB · URIR-8K | #Params: 11.8 -> 12.1 | Hurt |
Surprising findings
- Restormer's rain-accumulation PSNR (23.84 dB) is lower than the older CNN JORDER-E (23.76) only marginally, and all methods score lowest on rain accumulation, indicating it is the hardest pattern.
Limitations
- Rain patterns are synthetic and restricted to driving scenes, so real-world generalization is not directly validated on URIR-8K.
- Object detection evaluation relies on YOLOv5 pretrained models rather than detectors trained on the rain-free data.
Experimental setup
Datasets: URIR-8K Rain200L Rain200H UAV-Rain1K SPA-Data BDD100K
Benchmarks: URIR-8K Rain200L Rain200H UAV-Rain1K SPA-Data
Metrics: PSNR SSIM mAP50 FLOPs parameters
Baselines: LPNet JORDER-E RCDNet SPDNet IDT Restormer DRSformer PromptIR MambaIR Uformer
Builds on: MambaIR VMamba Vim PromptIR FGCM
Closest papers by meaning
- UniRain: Unified Image Deraining with RAG-based Dataset Distillation and Multi-objective Reweighted Optimization (CVPR 2026 · similarity 0.89)
- Harnessing Joint Rain-/Detail-aware Representations to Eliminate Intricate Rains (ICLR 2024 · similarity 0.88)
- Semantic Segmentation on Raindrop Degraded Images Using Two-Stage Dual Teacher-Student Learning (AAAI 2025 · similarity 0.88)
- AWRaCLe: All-Weather Image Restoration Using Visual In-Context Learning (AAAI 2025 · similarity 0.88)
- Bidirectional Multi-Scale Implicit Neural Representations for Image Deraining (CVPR 2024 · similarity 0.87)
- Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining (CVPR 2025 · similarity 0.87)
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.