Research map / Image generation / Image restoration
MODEM: A Morton-Order Degradation Estimation Mechanism for Adverse Weather Image Recovery
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 MODEM, a Mamba-style all-in-one adverse weather restoration network using Morton-order 2D selective scan plus a dual degradation estimator (global descriptor and spatially adaptive kernel) that modulates the scan via DAFM and DSAM, with two-stage training.
Problem
Restoring images degraded by diverse, spatially heterogeneous adverse weather (rain, haze, snow, raindrops) with a single unified model, which requires estimating spatially variant degradation.
Approach
Scan image features in Morton (Z-order) sequence within a selective state-space model, and condition the SSM via a dual degradation estimator: a global descriptor modulates features (FiLM-style) and a spatially adaptive kernel modulates attention; a teacher DDEM with ground truth supervises a student DDEM that sees only the degraded input.
Key results
- Average PSNR 34.18 dB and SSIM 0.9452 across Snow100K-S/L, Outdoor-Rain and RainDrop, versus 33.68 dB for Histoformer (the prior best).
- Outdoor-Rain PSNR 33.10 dB, a 1.02 dB gain over Histoformer's 32.08 dB.
- On RainDrop, PSNR 33.01 dB vs. Histoformer 33.06 dB, so MODEM is slightly below Histoformer on that one benchmark.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MODEM instead of Histoformer | Unified adverse-weather restoration · Snow100K-S | PSNR: 37.41 -> 38.08 | Helped |
| Use MODEM instead of Histoformer | Unified adverse-weather restoration · Snow100K-L | PSNR: 32.16 -> 32.52 | Helped |
| Use MODEM instead of Histoformer | Unified adverse-weather restoration · Outdoor | PSNR: 32.08 -> 33.10 | Helped |
| Use MODEM instead of Histoformer | Unified adverse-weather restoration · RainDrop | PSNR: 33.06 -> 33.01 | Hurt |
| Add Morton-order scan (ablation) | MODEM · Snow100K-S, Snow100K-L, Outdoor, RainDrop | PSNR: 38.03 -> 38.08; 32.33 -> 32.52; 32.89 -> 33.10; 32.69 -> 33.01 | Helped |
| Add DDEM (ablation) | MODEM · Snow100K-S, Snow100K-L, Outdoor, RainDrop | PSNR: 37.61 -> 38.08; 32.12 -> 32.52; 32.37 -> 33.10; 32.38 -> 33.01 | Helped |
| Add DSAM (ablation) | MODEM · Snow100K-S, Snow100K-L, Outdoor, RainDrop | PSNR: 38.15 -> 38.08; 32.59 -> 32.52; 32.77 -> 33.10; 32.72 -> 33.01 | Mixed |
| Add DAFM (ablation) | MODEM · Snow100K-S, Snow100K-L, Outdoor, RainDrop | PSNR: 37.43 -> 38.08; 32.07 -> 32.52; 32.19 -> 33.10; 32.62 -> 33.01 | Helped |
| Add correlation loss (ablation) | MODEM training · Snow100K-L | PSNR: 32.30 -> 32.52 | Helped |
| Add KL loss (ablation) | MODEM training · Snow100K-L | PSNR: 32.12 -> 32.52 | Helped |
| Use KL loss instead of the replaced KL-loss variant (ablation) | MODEM training · Snow100K-L | PSNR: 31.85 -> 32.52 | Helped |
| Use Morton scan instead of Hilbert scan (ablation) | MODEM · Snow100K-L | PSNR: 32.46 -> 32.52 | Helped |
Surprising findings
- Removing DSAM slightly improves some snow metrics (e.g., Snow100K-L PSNR 32.59 vs 32.52 with DSAM) while hurting raindrop removal.
Limitations
- Struggles with extremely large, high-contrast snowflakes that are underrepresented in training data.
- Larger parameter count (19.96M) than some Mamba baselines such as FreqMamba (8.93M).
Experimental setup
Datasets: Snow100K Raindrop Outdoor-Rain
Benchmarks: Snow100K-S Snow100K-L Outdoor-Rain RainDrop Test1 Snow100K-Real NTURain RESIDE WeatherStream
Metrics: PSNR SSIM LPIPS Q-Align MUSIQ
Baselines: All-in-One TransWeather WGWSNet WeatherDiff AWRCP Histoformer MambaIR FreqMamba VmambaIR ConvIR
Builds on: Mamba MambaIR Histoformer LC-Mamba
Closest papers by meaning
- Continuous Adverse Weather Removal via Degradation-Aware Distillation (CVPR 2025 · similarity 0.88)
- Real-World Adverse Weather Image Restoration via Dual-Level Reinforcement Learning with High-Quality Cold Start (NeurIPS 2025 · similarity 0.86)
- AWRaCLe: All-Weather Image Restoration Using Visual In-Context Learning (AAAI 2025 · similarity 0.85)
- GenDeg: Diffusion-based Degradation Synthesis for Generalizable All-In-One Image Restoration (CVPR 2025 · similarity 0.84)
- Multi-axis Prompt and Multi-dimension Fusion Network for All-in-one Weather-degraded Image Restoration (AAAI 2025 · similarity 0.84)
- FMRNet: Image Deraining via Frequency Mutual Revision (AAAI 2024 · similarity 0.84)
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.