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

Source: arXiv - 2601.23276v1

Overview

Astronomers constantly battle noisy CCD images that obscure faint celestial objects. This paper introduces a physics‑based noise synthesis framework that can generate realistic noisy‑clean image pairs for training deep‑learning denoisers—something that has been a major bottleneck for the community. By grounding the synthesis in the actual physical processes that create CCD noise, the authors make the resulting models both more trustworthy and easier to integrate into existing scientific pipelines.

Key Contributions

• Comprehensive CCD noise model covering photon shot noise, photo‑response non‑uniformity (PRNU), dark current, readout noise, and localized outliers (cosmic‑ray hits, hot pixels).
• Data‑generation pipeline that turns a high‑SNR “base” image (obtained by averaging many unregistered exposures) into unlimited paired noisy/clean frames for supervised learning.
• Real‑world multi‑band dataset captured with twin ground‑based telescopes, including raw frames, calibrated pipeline outputs, calibration frames (bias, dark, flat), and high‑SNR stacked bases.
• Demonstration of state‑of‑the‑art denoisers trained on the synthetic pairs, showing significant PSNR/SSIM gains over traditional calibration‑only methods on both simulated and real telescope data.
• Open‑source release of the noise synthesis code and the dataset, enabling reproducibility and rapid experimentation.

Methodology

1. High‑SNR Base Construction – The authors collect dozens of short exposures of the same sky region. By averaging them (without registration), random noise averages out while the underlying astronomical signal remains, yielding a clean “base” image.
2. Physics‑Based Noise Synthesis – Starting from the base, the pipeline injects noise components in the order they occur on a CCD:
   • Photon shot noise (Poisson‑distributed per pixel).
   • PRNU (pixel‑wise gain variations modeled as a low‑frequency multiplicative map).
   • Dark current (temperature‑dependent thermal electrons, modeled as a Poisson process plus a fixed pattern).
   • Readout noise (Gaussian electronic noise added after charge‑to‑voltage conversion).
   • Outliers (randomly placed high‑intensity spikes for cosmic rays and hot pixels).
     Each component is parameterized using calibration frames (bias, dark, flat) captured alongside the science data, ensuring that the synthetic noise matches the actual instrument behavior.
3. Dataset Assembly – The synthetic noisy images are paired with their original high‑SNR bases, forming a massive supervised training set. The authors also provide the raw telescope frames and the pipeline‑processed calibrated images for real‑world evaluation.
4. Training & Evaluation – Standard CNN‑based denoisers (e.g., DnCNN, UNet) and newer transformer‑style models are trained on the synthetic pairs. Performance is measured on both synthetic test sets and the real telescope data, using PSNR, SSIM, and astrophysical metrics such as source detection completeness.

Results & Findings

• Quantitative Gains – Denoisers trained on the physics‑based synthetic data achieve 3–5 dB higher PSNR and 10–15 % higher SSIM compared to traditional bias/dark/flat calibrated images on the real test set.
• Improved Source Recovery – The number of faint stars recovered at a 5σ detection threshold increased by ≈20 %, demonstrating that the denoising preserves astrophysical signal rather than smoothing it away.
• Outlier Handling – Incorporating cosmic‑ray and hot‑pixel models prevents the network from “learning” to erase genuine high‑intensity events, reducing false‑negative detections.
• Cross‑Instrument Generalization – Models trained on data from one telescope transferred well to a second, similarly equipped telescope, confirming that the physics‑based synthesis captures instrument‑agnostic noise characteristics.

Practical Implications

• Accelerated Pipeline Development – Researchers can now generate virtually unlimited training data for any CCD instrument simply by capturing a modest set of calibration frames and a few short exposures.
• Better Real‑Time Denoising – Deployable CNN/Transformer models can run on observatory GPUs, delivering near‑real‑time cleaned frames that aid in transient detection (e.g., supernovae, gravitational‑wave counterparts).
• Cost Savings – Reducing the need for long exposure stacking saves telescope time, allowing more sky coverage or deeper surveys within the same allocation.
• Standardization & Reproducibility – Because the noise model is rooted in measurable physical parameters, the denoising step can be audited and reproduced, a critical requirement for scientific publications.
• Cross‑Domain Adoption – The same framework can be adapted to other imaging sensors (CMOS, EMCCD) used in microscopy or remote sensing, where paired noisy‑clean data are also scarce.

Limitations & Future Work

• Model Assumptions – The current synthesis assumes stationary noise statistics across the detector; spatially varying temperature gradients or time‑dependent gain drift are not modeled.
• Outlier Complexity – Cosmic‑ray tracks can have complex shapes; the simplified spike model may not capture all morphological nuances, potentially limiting performance on heavily contaminated frames.
• Extension to Spectroscopy – The work focuses on broadband imaging; adapting the pipeline to spectroscopic CCD data (with wavelength‑dependent noise) remains an open challenge.
• Self‑Supervised Alternatives – Future research could explore combining the physics‑based synthetic pairs with self‑supervised denoising (e.g., Noise2Void) to further reduce reliance on calibration data.

---

The authors have released the code and dataset under an open license, making it straightforward for developers to experiment, benchmark, and integrate physics‑aware denoising into their own astronomical imaging workflows.

Authors

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

Paper Information

• arXiv ID: 2601.23276v1
• Categories: astro-ph.IM, cs.CV, cs.LG
• Published: January 30, 2026
• PDF: Download PDF