ASTRA-SR: Blind Single-Frame Atmospheric Turbulence Restoration for Planetary Imaging

Forum topic · 小凯 · 2026-09-24

Summary

ASTRA-SR is a blind single-frame image restoration framework addressing the joint problem of denoising, deblurring, and super-resolution in ground-based planetary imaging, which suffers from atmospheric turbulence, sensor noise, and limited sampling. The model is trained on a physics-grounded synthetic dataset: high-dynamic-range spacecraft RAW observations serve as clean sources, while paired low-resolution inputs are synthesized using measured layer-integrated turbulence strengths, propagated moving phase screens, exposure-averaged spatially varying PSFs, and sensor noise. The pipeline first estimates a noise-suppressed but blur-retaining LR image, then restores spatial structure via multiscale processing, and finally reconstructs high-resolution detail through serial spatial-amplitude refinement. ASTRA-SR achieves a 0.49 dB foreground PSNR improvement over the strongest baseline approaches. Paper by Xining Ge, Ziteng Cui, and Shuhong Liu, posted on arXiv (2609.26731).

论文概要

Field: Computer Vision Authors: Xining Ge, Ziteng Cui, Shuhong Liu Published: 2026-09-22 arXiv: 2609.26731

Abstract

Ground-based planetary imaging suffers from atmospheric turbulence, sensor noise, and limited sampling, making restoration a joint denoising, deblurring, and super-resolution problem. The authors present ASTRA-SR, a blind single-frame restoration framework trained on a physics-grounded synthetic dataset.

Dataset Construction

High-dynamic-range spacecraft RAW observations serve as clean source images.
Paired low-resolution (LR) inputs are synthesized using:
Measured layer-integrated turbulence strengths
Propagated moving phase screens
Exposure-averaged spatially varying PSFs
Sensor noise modeling

Method Pipeline

1. Estimate a noise-suppressed but blur-retaining LR image. 2. Restore spatial structure through multiscale processing. 3. Reconstruct high-resolution (HR) detail with serial spatial-amplitude refinement.

Results

ASTRA-SR yields a 0.49 dB foreground PSNR gain over the strongest baseline approaches.

---

*Auto-collected on 2026-09-24*