Research map / Image generation / Image restoration
Diffeomorphic Template Registration for Atmospheric Turbulence Mitigation
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 training-free method that registers atmospheric-turbulence frames to a latent template without explicitly initializing it, by averaging optical flows from a reference frame (justified by the Central Limit Theorem) and inverting them with a novel flow inversion algorithm.
Problem
Recovering the latent irradiance (template) from a sequence of images corrupted by atmospheric turbulence, where existing heuristic template initialization (e.g., averaging) yields blurry references that corrupt deformation estimates.
Approach
Select one frame as reference and estimate the template's deformation as the inverse of the mean of optical flows from that frame to all others (CLT-based), with an interpolation-based flow inversion so frames register to the template without explicitly constructing it.
Key results
- On HeatChamber, Ours-template reaches 20.30 PSNR / 0.780 SSIM, beating NDIR (20.08 / 0.763) without downstream processing; Ours-lucky reaches 20.74 / 0.789 vs TurbRecon 20.62 / 0.787.
- On CLEAR-sim (1024x1024), Ours-lucky achieves 25.95 PSNR / 0.906 SSIM vs TurbRecon 24.73 / 0.897 and NDIR 24.13 / 0.871.
- Processing 100 frames at 220x220 takes 17 seconds on a laptop CPU (31 seconds with ADMM deconvolution), versus about 30 minutes for NDIR on 10-minute sequences.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace NDIR's reference template with diffeomorphic template registration | HeatChamber multi-frame reconstruction · HeatChamber | SSIM: 0.763 -> 0.780 | Helped |
| Replace TurbRecon's space-time non-local averaging with diffeomorphic template registration | TurbRecon lucky image fusion pipeline · HeatChamber | PSNR: 20.62 -> 20.74 | Helped |
| Replace TurbRecon's space-time non-local averaging with diffeomorphic template registration | TurbRecon lucky image fusion pipeline · HeatChamber | SSIM: 0.787 -> 0.789 | Helped |
| Use diffeomorphic template registration with lucky image fusion instead of TurbRecon | Atmospheric turbulence reconstruction · CLEAR-sim | PSNR: 24.73 -> 25.95 | Helped |
| Use diffeomorphic template registration with lucky image fusion instead of TurbRecon | Atmospheric turbulence reconstruction · CLEAR-sim | SSIM: 0.897 -> 0.906 | Helped |
| Use the first frame rather than a selected sharp or blurry frame as key frame (ablation) | Diffeomorphic template registration · HeatChamber | SSIM: 0.770; 0.771; 0.771 | No clear effect |
| Add lucky image fusion to the registered template (ablation) | Ours-template · HeatChamber | PSNR: 20.30 -> 20.74 | Helped |
Surprising findings
- Using the first frame as reference performs as well as the sharpest (lucky) frame, with SSIM 0.771 vs 0.770 for the lucky frame in the frame-selection study.
- Single-image deep methods reach competitive PSNR but have low SSIM because they cannot model deformation.
Limitations
- Runtime is dominated by optical flow estimation, and the method is limited by generic optical flow not tailored to turbulence.
- Does not handle dynamic scenes or moving objects; occlusions violate diffeomorphic assumptions.
- Relies on assumptions (zero-mean deformation, many frames) for the CLT-based averaging.
Experimental setup
Datasets: HeatChamber CLEAR-sim HotAir CodedTarget
Benchmarks: HeatChamber CLEAR-sim
Metrics: PSNR SSIM runtime
Baselines: TRDN Uformer Restormer TurbNet PiRN CLEAR NDIR TurbRecon NDL
Builds on: TurbRecon Horn-Schunck optical flow SMURF ADMM blind deconvolution
Closest papers by meaning
- Spatio-Temporal Turbulence Mitigation: A Translational Perspective (CVPR 2024 · similarity 0.85)
- Temporally Consistent Atmospheric Turbulence Mitigation with Neural Representations (NeurIPS 2024 · similarity 0.83)
- Turb-Seg-Res: A Segment-then-Restore Pipeline for Dynamic Videos with Atmospheric Turbulence (CVPR 2024 · similarity 0.83)
- Learning Phase Distortion with Selective State Space Models for Video Turbulence Mitigation (CVPR 2025 · similarity 0.83)
- PlaNet: Learning to Mitigate Atmospheric Turbulence in Planetary Images (AAAI 2025 · similarity 0.82)
- STGC-NeRF: Spatial-Temporal Geometric Consistency for LiDAR Neural Radiance Fields in Dynamic Scenes (AAAI 2025 · similarity 0.82)
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.