Research map / Image generation / Image restoration
PlaNet: Learning to Mitigate Atmospheric Turbulence in Planetary Images
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 a vertical distance-aware turbulence simulator for planetary images and PlaNet, a permutation-invariant network accepting arbitrary frame counts with edge-based supervision from the average frame, outperforming existing turbulence mitigation methods.
Problem
Mitigating atmospheric turbulence in planetary images, where degradation patterns differ from terrestrial images, the number of input frames is arbitrary, and dark backgrounds provide little prior information.
Approach
A pixel-level turbulence simulator split into three atmospheric layers with layer-wise D/r0 configuration to match planetary degradation, plus a U-Net network with output-by-group permutation-invariant aggregation over any number of frames and multi-scale Laplacian edge supervision from the average frame.
Key results
- On synthetic data PlaNet reaches 27.78 PSNR / 0.901 SSIM, versus 26.48 / 0.887 for TurbNet and 24.41 / 0.836 for AutoStakkert.
- Removing edge-based supervision drops PSNR to 27.25; removing OBG (average pooling) drops it to 25.90.
- The simulator generates a 512x512 planetary image in 9 seconds on a 1080Ti GPU, hundreds of times faster than photon simulators.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PlaNet instead of TurbNet | Planetary image turbulence mitigation · Synthetic data | PSNR: 26.48 -> 27.78 | Helped |
| Use PlaNet instead of BIPNet | Planetary image turbulence mitigation · Synthetic data | SSIM: 0.898 -> 0.901 | Helped |
| Add edge-based supervision (ablation) | PlaNet · Synthetic data | PSNR: 27.25 -> 27.78 | Helped |
| Use OBI for edge supervision rather than applying it after feature-map aggregation (ablation) | PlaNet · Synthetic data | PSNR: 27.21 -> 27.78 | Helped |
| Use OBG instead of average pooling (ablation) | PlaNet · Synthetic data | PSNR: 25.90 -> 27.78 | Helped |
| Use multi-scale Ldec rather than calculating it only at level j = 1 (ablation) | PlaNet · Synthetic data | PSNR: 26.40 -> 27.78 | Helped |
| Aggregate features rather than frames before the backbone (ablation) | PlaNet · Synthetic data | PSNR: 23.37 -> 27.78 | Helped |
Surprising findings
- Frame-level aggregation before the backbone (instead of feature-level aggregation) drops PSNR to 23.37, below several baselines.
Limitations
- Real-world data only covers planets, not other space objects.
- Edge-based supervision may be less effective when backgrounds are not scarce or edges are unclear.
Experimental setup
Datasets: synthetic planetary image sequences from 27 NASA 3D solar system models
Benchmarks: synthetic planetary test set (5 3D models, 60 sequences) real planetary images captured with Celestron C11 and internet samples
Metrics: PSNR SSIM
Baselines: AutoStakkert TurbNet NIDR BIPNet BasicVSR++
Builds on: TurbNet BasicVSR++ Mao et al. 2021 P2S turbulence simulation U-Net CBAM
Closest papers by meaning
- Spatio-Temporal Turbulence Mitigation: A Translational Perspective (CVPR 2024 · similarity 0.86)
- Learning Phase Distortion with Selective State Space Models for Video Turbulence Mitigation (CVPR 2025 · similarity 0.84)
- NB-GTR: Narrow-Band Guided Turbulence Removal (CVPR 2024 · similarity 0.84)
- Temporally Consistent Atmospheric Turbulence Mitigation with Neural Representations (NeurIPS 2024 · similarity 0.83)
- Diffeomorphic Template Registration for Atmospheric Turbulence Mitigation (CVPR 2024 · similarity 0.82)
- Turb-Seg-Res: A Segment-then-Restore Pipeline for Dynamic Videos with Atmospheric Turbulence (CVPR 2024 · similarity 0.81)
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.