Learning the Night Sky with Deep Generative Priors

TL;DR

The paper tackles the problem of recovering sharp, photometry-ready images from blurred, noisy ground-based astronomical exposures by proposing an unsupervised multi-frame deconvolution framework that leverages deep generative priors. It encodes the latent sky image $x$ as a function of multiple exposures using a CNN-based network $F_{\theta}$ and reconstructs exposure-specific outputs via convolution with varying PSFs $f^{(t)}$, within a Bayesian MAP setup under a Gaussian likelihood $p(y|x,f)$. A robust training objective based on a scaled Huber loss, nonnegativity via ReLU, and a CNN architecture with parallel depthwise convolutions enables cross-frame regularization without supervision. Demonstrations on 4K×4K Hyper Suprime-Cam data show improved restorations over naive co-adds, yielding a physically meaningful $\hat{x}$ suitable for photometry and surface-detail improvement of sources; the work lays the groundwork for extensions to multi-band data and super-resolved latent images for sub-pixel photometry in upcoming surveys like the Rubin Observatory.

Abstract

Recovering sharper images from blurred observations, referred to as deconvolution, is an ill-posed problem where classical approaches often produce unsatisfactory results. In ground-based astronomy, combining multiple exposures to achieve images with higher signal-to-noise ratios is complicated by the variation of point-spread functions across exposures due to atmospheric effects. We develop an unsupervised multi-frame method for denoising, deblurring, and coadding images inspired by deep generative priors. We use a carefully chosen convolutional neural network architecture that combines information from multiple observations, regularizes the joint likelihood over these observations, and allows us to impose desired constraints, such as non-negativity of pixel values in the sharp, restored image. With an eye towards the Rubin Observatory, we analyze 4K by 4K Hyper Suprime-Cam exposures and obtain preliminary results which yield promising restored images and extracted source lists.

Figures (2)

• Figure 1: Network architecture. We extract and combine multi-scale information from the exposures via various convolutional layers. All input exposures first pass through several depth-wise convolution layers in parallel. The output channels are concatenated, and a pointwise convolution completes the "encoder" part of our network to produce the latent image $x$. Desired constraints, such as non-negativity of pixel values in $x$, are enforced by applying a ReLU activation. We then "decode" $x$ via a final 2D convolutional layer in order to produce the reconstructions.
• Figure 2: Comparison: Selected cutouts from a raw exposure (left) vs. the "sample mean" co-add (middle) vs. a restored image $\Hat{x}$ from our approach (right). Our method produces a physically meaningful restored latent image of the night sky which is suitable for photometry. Pixels in the sky background have zero intensity, with only minor speckles of noise appearing around bright objects, unlike the co-add where noise is uniformly present across the image. Moreover, our method successfully deblurs a wide array of sources, resulting in e.g. galaxy shapes being visibly more well-defined, and large blurry stars appearing as well resolved point sources.