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

Shuhong Liu, Xining Ge, Ziying Gu, Lin Gu, Ziteng Cui, Xuangeng Chu, Jun Liu, Dong Li, Tatsuya Harada · 2026

Read Denoising the Deep Sky: Physics-Based CCD Noise Formation for Astronomical Imaging by Shuhong Liu, Xining Ge, Ziying Gu, Lin Gu, Ziteng Cui, Xuangeng Chu, Jun Liu, Dong Li, Tatsuya Harada (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-23276.txt.

Introduction

Shuhong Liu and eight co-authors model CCD noise from physics to clean deep-sky images. Sky imaging stays noise-limited under real observing conditions. Standard pipelines remove structured flaws, but random noise in the sensor is mostly left unresolved.

Learning-based cleaning is promising, yet paired training data are scarce and science needs interpretable models. Their physics-based making framework generates realistic CCD noise for training. Faithful noise models let cleaners learn the sensor's true behavior instead of generic blur. Physically interpretable, reproducible noise models are a stated requirement, since scientific imaging cannot rest on black-box flaws.

Catalog record

CREATORShuhong Liu, Xining Ge, Ziying Gu, Lin Gu, Ziteng Cui, Xuangeng Chu, Jun Liu, Dong Li, Tatsuya Harada
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE51 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-23276

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Denoising the Deep Sky: Physics-Based CCD Noise Formation for Astronomical Imaging Shuhong Liu, Xining Ge, Ziying Gu, Lin Gu, Ziteng Cui, Xuangeng Chu, Jun Liu, Dong Li, Tatsuya Harada 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 abundent 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. 1 Introduction Modern imaging systems devote substantial effort to suppressing sensor noise through advances in both hardware design and computational methods. Nevertheless, digital images remain affected by fluctuations that are intrinsic to the process of photon detection and electronic readout, as well as by residual artifacts introduced by the sensor architecture [ 45 ] . Reducing these noise sources while faithfully preserving signal fidelity has therefore become a central objective for computational imaging. For consumer photography, particularly in extreme low-light conditions, learning-based denoising has advanced substantially. Learning to See in the Dark [ 9 ] demonstrated the efficacy of learning in linear RAW space to reconstruct high signal-to-noise ratio (SNR) images from noisy measurements. To mitigate the scarcity in acquiring strictly paired data, subsequent research has focused on developing sophisticated noise modeling and synthesis techniques [ 67 , 61 , 42 , 29 , 16 ] . Within this setting, physics-based models [ 60 , 8 , 33 ] , implicit modeling [ 15 , 64 ] , and non-parametric approaches [ 10 , 43 ] have been proposed to synthesize camera-dependent noisy RAW samples at scale, tailored to the ISO-aware noise characteristics of the modern Complementary Metal-Oxide-Semiconductor (CMOS) sensors that are prevalent in consumer cameras. 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. Scientific imaging, particularly in astronomy, presents a different landscape where Charge-Coupled Device (CCD) sensors remain the dominant technology [ 23 ] . The CDD architecture differs fundamentally from CMOS cameras. Instead of embedding amplifiers and converters at the pixel level, CCDs transfer charge serially to a small set of output amplifiers [ 28 ] . This readout process introduces distinctive noise signatures [ 14 ] , including segment-specific bias and gain variation [ 31 ] , charge transfer inefficiency [ 37 ] , clock-induced charge [ 27 ] , and bias drift and striping [ 21 ] . At longer wavelengths, back-illuminated CCDs exhibit interference fringes or etaloning [ 23 ] , caused by internal reflections within the silicon substrate. Unlike consumer CMOS sensors, science CCDs record in a single spectral band defined by filters rather than through a mosaic color pattern, which alters the statistical structure of the data. Astronomical CCDs are also exposed to unique observational effects, including cosmic-ray strikes [ 57 ] and persistent hot pixels [ 39 ] that introduce localized outliers. Unlike consumer CMOS imaging, where noise can often be reduced by simply extending the exposure time, astronomical observations are fundamentally constrained by atmospheric effects and target dynamics. Ground-based telescopes are subject to turbulence-induced wavefront aberrations [ 18 ] , scintillation [ 49 ] , intrinsic source variability [ 63 ] , and Earth’s rotation [ 5 ] , all of which evolve on timescales that make long integrations for obtaining high-SNR images impractical. As a result, CCD observations are typically acquired as a sequence of shorter exposures that are individually calibrated and then registered and stacked [ 19 ] . While stacking improves SNR and enables the rejection of outliers, it also faces notable limitations. Residual offsets from Earth’s rotation and imperfect telescope tracking can prevent accurate co-registration [ 6 ] . Moreover, the strict alignment required for stacking is computationally intensive and not always feasible [ 54 ] , especially for faint or time-variable sources [ 3 ] , or for large-area survey observations where rapid coverage is prioritized [ 26 ] . Consequently, the scientific usability of individual exposures remains essential, while conventional calibration pipelines primarily suppress structured artifacts and leave stochastic noise largely unresolved [ 23 ] . To learn the denoising model for astronomical imaging and overcome the data sparsity commonly seen in scientific imaging, we propose a physics-based noise synthesis framework tailored to CCD noise formation. As illustrated in Figure 1 , our pipeline models the principal sources of CCD noise, including photon shot noise, per-amplifier bias offsets and gain non-uniformity, dark-current noise, and readout effects such as read noise and charge transfer inefficiency at the output amplifier. It further accounts for observational phenomena, including transient cosmic-ray hits and persistent hot pixels, which can noticeably affect raw exposures. A natural question is how to obtain the“clea” bases for synthesis, since truly noise-free references are impractical. We leverage common practice in modern telescopes that track targets with active guiding so the scene remains approximately fixed on the detector across sequential exposures. We directly average multiple unregistered observation frames, which inevitably broadens the astrophysical point-spread function due to residual drift and seeing variation, but substantially suppresses stochastic noise without introducing interpolation artifacts. Such averaged images are not scientifically usable per se; however, they provide abundant high-SNR bases that preserve correct large-scale background structure and instrument signatures. Our synthesis pipeline can generates realistic noisy counterparts from these bases, enabling the construction of paired datasets for supervised learning. To validate the framework, we further construct a real astronomical image dataset that contains paired raw noisy frames and instrument-pipeline calibrated frames collected from two twin ground-based telescope observations for real-world evaluation. Our main contributions can be summarized as follows: • We propose a physics-based noise formulation model tailored to astronomical CCD imaging, capable of synthesizing realistic noisy RAW data that closely matches the statistical properties of real observations. • We overcome the scarcity of clean data by presenting a practical method that averages unregistered multi-frame observations into high-SNR bases. These bases provide effective inputs for physics-based noise synthesis with minimal preparation effort. • We construct real-world multi-band astronomical imaging datasets from ground telescope observations of outer solar system targets, providing noisy RAW images, calibration-processed references, calibration frames, and stacked high-SNR clean bases for noise synthesis and quantitative evaluation. 2 Related Work 2.1 Desnoising in Consumer CMOS Imaging Recent efforts in CMOS image denoising

Read the complete text