Denoising the Deep Sky: Physics-Based CCD Noise Formation for Astronomical Imaging

TL;DR

This work tackles the persistent challenge of stochastic CCD noise in astronomical imaging by introducing a physics-based noise formation model tailored to CCDs. It decomposes noise into signal-dependent and signal-independent components (including PRNU, photon-shot noise, dark current, fixed-pattern read noise, digitization, and impulsive outliers) and synthesizes realistic noisy RAW images from high-SNR bases formed by averaging unregistered frames. A U-Net denoiser is trained on these physics-based synthetic pairs, and evaluated on real MuSCAT data (MuSCAT-3) and zero-shot on MuSCAT-4, showing improved background fidelity (lower NMAD/STD) and photometric preservation compared with traditional and learning-based baselines. The approach offers interpretable, reproducible noise modeling that enables reliable supervised denoising for precision astronomy, with demonstrated cross-instrument generalization and a dedicated real-world multi-band dataset for evaluation. Limitations include reliance on calibrated proxies for ground truth and validation primarily within MuSCAT instruments, indicating scope for broader validation across more telescopes and sensors.

Abstract

Astronomical imaging remains noise-limited under practical observing constraints, while standard calibration pipelines mainly remove structured artifacts and leave stochastic noise largely unresolved. Learning-based denoising is promising, yet progress is hindered by scarce paired training data and the need for physically interpretable and reproducible models in scientific workflows. We propose a physics-based noise synthesis framework tailored to CCD noise formation. The pipeline models photon shot noise, photo-response non-uniformity, dark-current noise, readout effects, and localized outliers arising from cosmic-ray hits and hot pixels. To obtain low-noise inputs for synthesis, we average multiple unregistered exposures to produce high-SNR bases. Realistic noisy counterparts synthesized from these bases using our noise model enable the construction of abundant paired datasets for supervised learning. We further introduce a real-world dataset across multi-bands acquired with two twin ground-based telescopes, providing paired raw frames and instrument-pipeline calibrated frames, together with calibration data and stacked high-SNR bases for real-world evaluation.

Figures (10)

• Figure 1: Overview of denoising pipeline. Left: a CMOS sensor performs in-pixel conversion and shows a near-uniform pixel responsivity gain, while a scientific CCD sensor transfers pixel charge and exhibits spatially structured gain. Right: paired training images are limited since clean signals cannot be recovered by extending exposure or traditional calibration. We stack raw inputs to form a high-SNR baseline, synthesize realistic noisy raws using a physics-based noise model, and train the denoising network.
• Figure 2: Overview of noise formation in astronomical imaging. Unique telescope optics and CCD science cameras introduce noise characteristics that are distinct from commercial CMOS imaging. Zoom in the noise frames for better visibility.
• Figure 3: PRNU comparison between $G^{\raisebox{-0.60ex}{$\prime$}}$ and $R^{\raisebox{-0.60ex}{$\prime$}}$ bands. A captured sky flat frame (a) is typically expressed as the product of (b) the anti-fringing fixed pattern, (c) the low-frequency twilight sky background, and (d) dust-donut artifacts. PRNU is given by (b) $\times$ (d), where (c) is the external signal.
• Figure 4: Quantile--Quantile plots of fitting Gaussian models to dark frames residuals.
• Figure 5: Fixed-pattern noise and read noise within 97% sigma-clipp.