ASTRODIFFASTROPHOTOGRAPHY TURBULENCE MITIGATION VIA GENERATIVE MODELS

Joonyeoup Kim Yu Yuan Xingguang Zhang Xijun Wang Stanley Chan

Abstract

Photography is the cornerstone of modern astronomical and space research. However, most astronomical images captured by ground-based telescopes suffer from atmospheric turbulence, resulting in degraded imaging quality. While multi-frame strategies like lucky imaging can mitigate some effects, they involve intensive data acquisition and complex manual processing. In this paper, we propose AstroDiff, a generative restoration method that leverages both the high-quality generative priors and restoration capabilities of diffusion models to mitigate atmospheric turbulence. Extensive experiments demonstrate that AstroDiff outperforms existing state-of-the-art learning-based methods in astronomical image turbulence mitigation, providing higher perceptual quality and better structural fidelity under severe turbulence conditions.

Methods

Our training process is divided into two stages. In the first stage, we perform extensive pre-training separately for the generative prior branch and the restoration branch. In the second stage, we jointly combine these two models through a fusion module using Stochastic Gradient Langevin Dynamics inference.

Baseline Comparison

AstroDiff's performance comparison against 3 state-of-the-art turbulence mitigation models (DATUM, TMT, ESTRNN) on planetary images with strong synthetic turbulence

Moon - AstroDiff

Moon - DATUM

Moon - TMT

Moon - ESTRNN

Mercury - AstroDiff

Mercury - DATUM

Mercury - TMT

Mercury - ESTRNN

Jupiter - AstroDiff

Jupiter - DATUM

Jupiter - TMT

Jupiter - ESTRNN

← EnhancedOriginal →

Method	Low Turb		Medium Turb		High Turb
Method	PSNR ↑	LPIPS ↓	PSNR ↑	LPIPS ↓	PSNR ↑	LPIPS ↓
DATUM	31.09	0.11	32.12	0.09	30.44	0.12
TMT	29.57	0.13	30.03	0.13	30.49	0.11
ESTRNN	29.02	0.18	29.58	0.16	28.83	0.14
AstroDiff (ours)	29.29	0.07	30.22	0.08	30.86	0.07

Real-Image Results

Slide to compare the original and enhanced images of celestial bodies. Our method significantly improves the image quality while preserving important details.