AI breakthrough restores blurry space images without prior distortion data

A new study has shown that a U-Net, a type of artificial neural network, can independently restore blurry astronomical images without needing prior details about the distortion. This breakthrough offers a fresh approach to processing space observations, where traditional methods often struggle with incomplete or inaccurate data.

Researchers trained the U-Net using simulated astronomical data generated with the GalSim toolkit. They incorporated real-world references from the Zwicky Transient Facility (ZTF) and the Vera C. Rubin Observatory's Legacy Survey of Space and Time (LSST), both publicly available through their respective data portals. The model's performance improved as the training dataset grew, eventually reaching a saturation point where additional data yielded diminishing returns.

The U-Net outperformed classical deconvolution techniques like Tikhonov's method, especially in difficult conditions with unknown noise or blurring effects. Unlike traditional approaches, which depend on precise estimates of the point spread function (PSF) and noise levels, the U-Net operated as a standalone system, requiring no prior information about these factors.

Tests revealed the network's strong generalisation abilities, handling unseen noise and atmospheric distortion with consistent accuracy. The study also suggested the U-Net learns a geometry-adaptive harmonic basis, a behaviour aligned with recent mathematical theories on its inner workings. This adaptability mirrors patterns seen in denoising tasks, reinforcing its potential for solving complex inverse problems in astronomy.

The findings position the U-Net as a robust tool for astronomical image restoration, reducing reliance on error-prone assumptions about blurring or noise. Its ability to learn and adapt from data alone could streamline processing pipelines for future observatories, including the upcoming Vera C. Rubin Observatory's LSST survey.