AI breakthrough: U-Net restores blurred astronomical images without prior data

A new study has revealed how a U-Net, a type of artificial neural network, can independently restore blurred astronomical images without needing prior details about the blurring process. Traditional methods often struggle because they depend on accurate estimates of distortion and noise—information that is frequently hard to obtain.

Researchers trained the U-Net model using simulated astronomical data generated with the GalSim toolkit. The datasets included up to 40,000 images of size 48x48 from the COSMOS Real Galaxy Dataset, incorporating random transformations, optical and atmospheric blurring effects, and Gaussian white noise. No real telescope observations were used for validation.

The U-Net demonstrated strong performance, effectively handling unseen conditions such as varying noise levels and optical distortions. Unlike traditional deconvolution techniques, it worked as a standalone system, processing images end-to-end without prior knowledge of the point spread function or noise characteristics.

Tests showed the model consistently improved with more training data. It also outperformed the classical Tikhonov deconvolution method, especially in complex scenarios. The study suggests the U-Net learns a geometry-adaptive harmonic basis, similar to sparse representations seen in denoising tasks.

These findings align with recent mathematical research highlighting the adaptive learning strengths of U-Net architectures. The network's ability to generalise across different conditions makes it a promising tool for astronomical image processing and broader applications in solving ill-posed inverse problems.

The U-Net's success offers astronomers a more flexible approach to image restoration, reducing reliance on hard-to-obtain measurements. Its ability to adapt and improve with larger datasets could lead to broader use in fields where precise prior information is unavailable.