Research map / Continual learning / Online and replay-based
Continuous Adverse Weather Removal via Degradation-Aware Distillation
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 ILAWR, an incremental all-in-one adverse weather removal framework using Fourier-prior degradation-aware module and importance-guided multi-teacher distillation, without sample replay or parameter expansion.
Problem
Continual adverse weather removal with all-in-one restoration models suffers from catastrophic forgetting of earlier degradation types and limited degradation awareness, while existing incremental methods rely on replay caches, parameter expansion, or regularization that are memory- or compute-costl
Approach
A Degradation-Aware Module built on Fourier amplitude priors isolates degradation-specific channel features; a multilateral distillation scheme uses two teacher groups (DAM-based degradation teachers and previous restoration models for background), aggregated with KL-based importance weights (IGAM), so no old data is stored.
Key results
- Setting 1 (Haze→Rain→Snow): ILAWR average PSNR 31.32 dB vs CLAIO 30.70 dB and sequential fine-tuning 20.63 dB.
- Setting 2 (four weather types): ILAWR average PSNR 30.13 dB, about 0.85 dB above the best compared method CLAIO (29.28 dB).
- Setting 3 (real-world, low-light added): ILAWR average PSNR 26.78 dB vs WGWS 25.68 dB and CLAIO 25.62 dB; uses 4.95M parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DAM and IGAM (ablation) | ILAWR · Rain100H (Setting 1) | PSNR: 13.47 -> 28.91 | Helped |
| Add DAM (ablation) | ILAWR · Rain100H (Setting 1) | PSNR: 22.16 -> 28.91 | Helped |
| Add IGAM (ablation) | ILAWR · Rain100H (Setting 1) | PSNR: 22.37 -> 28.91 | Helped |
| Add Lteach1 and Lteach2 (ablation) | ILAWR · Rain100H (Setting 1) | PSNR: 13.47 -> 28.91 | Helped |
| Add Lteach1 (ablation) | ILAWR · Rain100H (Setting 1) | PSNR: 22.09 -> 28.91 | Helped |
| Add Lteach2 (ablation) | ILAWR · Rain100H (Setting 1) | PSNR: 22.45 -> 28.91 | Helped |
| Change weather-task order (ablation) | ILAWR · Setting 1 synthetic datasets | PSNR: 31.28 to 31.39 | No clear effect |
| Increase ζ excessively (ablation) | ILAWR · Setting 1 synthetic datasets | Average PSNR | Hurt |
| Increase λ excessively (ablation) | ILAWR · Setting 1 synthetic datasets | Average PSNR | Hurt |
| Use ILAWR instead of CLAIO | Incremental weather removal · Setting 1 average | PSNR: 30.70 -> 31.32 | Helped |
| Use ILAWR instead of CLAIO | Incremental weather removal · Setting 2 average | PSNR: 29.28 -> 30.13 | Helped |
| Use ILAWR instead of WGWS | Incremental weather removal · Setting 3 real datasets average | PSNR: 25.68 -> 26.78 | Helped |
Surprising findings
- Sequential fine-tuning on Snow100K retains 32.21 dB, higher than many incremental methods on that dataset, while forgetting heavily on RESIDE and Rain100H.
Limitations
- Relies on teacher models from previous sessions and per-session full model copies, so guidance pool grows with number of sessions.
- Evaluated on a single weather order family with a small number of sessions (3-4); ablations only in Setting 1.
Experimental setup
Datasets: RESIDE Rain100H Snow100K Raindrop REVIDE SPA+ RealSnow LOL-v2
Benchmarks: RESIDE Rain100H Snow100K Raindrop REVIDE SPA+ RealSnow LOL-v2
Metrics: PSNR SSIM parameter count inference time
Baselines: Sequential Fine-tuning EWC LwF MAS POD PIGWM AFC TransWeather MutiTS WGWS RAM CLAIO
Builds on: CLAIO PIGWM U-Net contrast regularization Learning without Forgetting
Closest papers by meaning
- Real-World Adverse Weather Image Restoration via Dual-Level Reinforcement Learning with High-Quality Cold Start (NeurIPS 2025 · similarity 0.89)
- MODEM: A Morton-Order Degradation Estimation Mechanism for Adverse Weather Image Recovery (NeurIPS 2025 · similarity 0.88)
- Language-driven All-in-one Adverse Weather Removal (CVPR 2024 · similarity 0.88)
- AWRaCLe: All-Weather Image Restoration Using Visual In-Context Learning (AAAI 2025 · similarity 0.88)
- Genuine Knowledge from Practice: Diffusion Test-Time Adaptation for Video Adverse Weather Removal (CVPR 2024 · similarity 0.87)
- Semantic Segmentation in Multiple Adverse Weather Conditions with Domain Knowledge Retention (AAAI 2024 · 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.