Research map / Image generation / Image restoration
Multi-axis Prompt and Multi-dimension Fusion Network for All-in-one Weather-degraded Image Restoration
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 MPMF-Net, an all-in-one traffic weather-degradation restoration network with multi-axis prompts (MPLB), multi-dimension feature fusion (MFIB), and coarse-scale degradation-free implicit neural representations (CDINR).
Problem
All-in-one removal of adverse weather degradations (haze, rain, snow, low light) from traffic images, where existing methods lose image quality and are slow at inference.
Approach
Learns prompts along three axis planes (hw, ch, cw) instead of only spatial or channel dimensions, fuses encoder-decoder features via height/width/channel mutual attention with learnable scaling, and uses coarse-scale implicit neural representations as self-prompts to normalize degradation levels.
Key results
- Averaged over six datasets, MPMF-Net improves PSNR over PromptIR by 0.576 dB and over DyNet by 0.684 dB.
- Ablation on the traffic all-in-one benchmark: CDINR adds 0.741 dB, MFIB adds 0.516 dB, and MPLB adds 0.718 dB over the baseline.
- MPMF-Net uses about 62.765% of PromptIR's inference time (0.0622 vs 0.0991) with 3.2393M parameters, the fewest among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CDINR to the baseline (ablation) | MPMF-Net · Six test datasets (average) | PSNR: +0.741 dB | Helped |
| Replace convolution fusion with MFIB (ablation) | MPMF-Net · Six test datasets (average) | PSNR: +0.516 dB | Helped |
| Add MPLB to the decoding stage (ablation) | MPMF-Net · Six test datasets (average) | PSNR: +0.718 dB | Helped |
| Replace simple addition with convolution fusion (ablation) | Feature fusion model · Six test datasets (average) | PSNR: 29.876 -> 30.244 | Helped |
| Replace convolution with SKFusion (ablation) | Feature fusion model · Six test datasets (average) | PSNR: +0.674 dB | Helped |
| Replace SKFusion with MFIB (ablation) | Feature fusion model · Six test datasets (average) | PSNR: +0.669 dB | Helped |
| Use MPLB instead of TransWeather prompt learning (ablation) | MPMF-Net · Six test datasets (average) | PSNR: +2.265 dB | Helped |
| Use MPLB instead of AIRFormer prompt learning (ablation) | MPMF-Net · Six test datasets (average) | PSNR: +1.757 dB | Helped |
| Use MPLB instead of PromptIR prompt learning (ablation) | MPMF-Net · Six test datasets (average) | PSNR: +0.833 dB | Helped |
| Use 1/2-scale rather than 1/4-scale input for CDINR (ablation) | MPMF-Net · Six test datasets (average) | PSNR: 31.587 -> 30.933 | Hurt |
| Use 1/8-scale rather than 1/4-scale input for CDINR (ablation) | MPMF-Net · Six test datasets (average) | PSNR: 31.587 -> 30.525 | Hurt |
| Use original-scale rather than 1/4-scale input for CDINR (ablation) | MPMF-Net · Six test datasets (average) | PSNR: 31.587 -> 29.580 | Hurt |
Surprising findings
- Image-scale ablation shows that representing clean images at the original scale in CDINR performs worst, with 1/4-scale inputs performing best.
Limitations
- Trained only on six weather degradation types, so it cannot handle complex conditions such as dust or back-light.
- Evaluated only on image reconstruction, not as a pre-processing step for downstream high-level vision tasks.
Experimental setup
Datasets: FoggyCityscapes RainCityscapes RSCityscapes SnowTrafficData LowLightTrafficData RainDS-syn
Benchmarks: FoggyCityscapes RainCityscapes RSCityscapes SnowTrafficData LowLightTrafficData RainDS-syn
Metrics: PSNR SSIM LPIPS inference time #Param
Baselines: All-in-One TransWeather TKL AirNet AIRFormer WeatherDiff PromptIR DyNet
Builds on: PromptIR Restormer AirNet TransWeather AIRFormer
Closest papers by meaning
- 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 0.86)
- Continuous Adverse Weather Removal via Degradation-Aware Distillation (CVPR 2025 · similarity 0.86)
- UP-Restorer: When Unrolling Meets Prompts for Unified Image Restoration (AAAI 2025 · similarity 0.85)
- Rethinking Multi-Scale Representations in Deep Deraining Transformer (AAAI 2024 · similarity 0.85)
- Language-driven All-in-one Adverse Weather Removal (CVPR 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.