🔭 astrophysics

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

This paper proposes a physics-based noise synthesis framework that models various CCD noise sources to generate abundant paired training data from stacked astronomical exposures, thereby enabling effective supervised learning for denoising deep-sky images while maintaining scientific accuracy.

Original authors: Shuhong Liu, Xining Ge, Ziying Gu, Quanfeng Xu, Lin Gu, Ziteng Cui, Xuangeng Chu, Jun Liu, Dong Li, Tatsuya Harada

Published 2026-03-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shuhong Liu, Xining Ge, Ziying Gu, Quanfeng Xu, Lin Gu, Ziteng Cui, Xuangeng Chu, Jun Liu, Dong Li, Tatsuya Harada

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to take a photograph of a tiny, distant firefly in a pitch-black forest. But there's a catch: your camera is old, it's freezing cold, and every time you press the shutter, a few random snowflakes land on the lens, and the camera's internal electronics add a static "hiss" to the picture.

This is exactly what astronomers face when they try to take pictures of deep space. The stars are faint, the cameras (called CCDs) are complex, and the resulting images are full of "noise"—graininess, weird streaks, and random bright dots that aren't actually stars.

Here is the story of how this paper solves that problem, explained simply.

The Problem: The "Noisy" Telescope

In regular photography (like your phone), if an image is too dark, you can just hold the camera steady for a long time to let more light in. But in astronomy, you can't just "hold still" for hours. The Earth is spinning, the atmosphere is wiggling (like heat haze), and the stars themselves might be flickering.

So, astronomers take hundreds of short, quick snapshots. They try to stack them together to make one clear picture. But even after stacking, the images are still grainy.

The Big Hurdle: To teach a computer (AI) to clean up these images, you usually need to show it a "dirty" picture and the "clean" version of the same picture. But in space, there is no such thing as a "clean" picture. You can't take a perfect photo of a distant galaxy to use as a reference. Without a clean reference, the AI doesn't know what it's supposed to look like.

The Solution: The "Physics Chef"

The authors of this paper realized that instead of waiting for a perfect photo, they could fake the noise using a recipe based on real physics.

Think of it like a chef trying to teach a student how to fix a soup that tastes too salty. Instead of finding a perfect, unsalted soup (which doesn't exist), the chef takes a bowl of perfect, clear water (a high-quality, stacked image) and deliberately adds salt, pepper, and a little bit of dirt to it, following a strict recipe.

The Clean Base: They take many short, blurry photos of the same star and stack them together. This creates a "super-image" that is very clear but still a bit fuzzy. It's not perfect, but it's the best "clean" version they can get.
The Noise Recipe: They built a computer model that knows exactly how their specific telescope camera works. They know:
- Photon Shot Noise: How light arrives in random packets (like raindrops hitting a roof).
- PRNU: How some pixels on the camera are slightly "dumber" or "brighter" than others (like a patch of uneven grass).
- Dark Current: How the camera gets "hot" and creates fake signals even in the dark (like a warm engine humming).
- Cosmic Rays: Random high-energy particles from space that hit the sensor and leave bright white dots (like a fly buzzing across the lens).
The Training: They take their "super-image" and use their recipe to add all these noises back in, creating thousands of realistic "dirty" photos. Now, they have a perfect pair: the "clean" base and the "synthesized dirty" version. They feed this to the AI to teach it how to clean up the mess.

The Result: A Magic Eraser for Space

Once the AI is trained on this "physics-based" recipe, they tested it on real telescope data.

The Old Way: Traditional methods either smoothed out the image so much that the stars looked like blurry blobs, or they left too much grain.
The New Way: Their AI cleaned up the noise but kept the stars sharp and the colors accurate. It was like using a magic eraser that removed the dust but didn't smudge the drawing.

Why This Matters

This is a game-changer for astronomy because:

It's Scientific: It doesn't just make the picture look pretty; it preserves the exact brightness of the stars. This is crucial for measuring how far away stars are or if they are changing size.
It's General: They tested it on two different telescopes, and it worked on both without needing to be retrained. It's like a universal translator that understands the "language" of noise for different cameras.
It Solves the Data Scarcity: They didn't need millions of perfect photos to train the AI. They just needed a good understanding of physics and a few calibration frames.

In a nutshell: The authors figured out how to simulate the "mess" of a telescope camera so perfectly that they could train an AI to clean up real space photos, revealing the universe with unprecedented clarity.

1. Problem Statement

Astronomical imaging faces a unique challenge: while standard calibration pipelines (e.g., bias, dark, flat-field correction) remove structured artifacts, they leave stochastic noise (photon shot noise, read noise, dark current) largely unresolved. This noise limits the sensitivity required to detect faint celestial objects.

Data Scarcity: Unlike consumer photography, obtaining "clean" ground-truth images for supervised learning in astronomy is impractical. Extending exposure times is often impossible due to atmospheric turbulence, target dynamics, and telescope rotation.
Model Mismatch: Existing deep learning denoising methods are primarily designed for consumer CMOS sensors. These models fail to generalize to scientific CCD sensors because:
- CCDs use serial charge transfer, introducing unique noise signatures (e.g., column-wise banding, charge transfer inefficiency) absent in CMOS.
- Astronomical data lacks the Bayer color mosaic, operating in single spectral bands with different statistical structures.
- Astronomical observations suffer from unique outliers like cosmic-ray hits and persistent hot pixels.
Training Data Gap: There is a lack of large-scale, paired datasets (clean vs. noisy) for training deep learning models in astronomy.

2. Methodology

The authors propose a physics-based noise synthesis framework to generate realistic paired training data without requiring clean ground truth. The pipeline consists of three main stages:

A. High-SNR Base Generation

To create a "clean" reference for synthesis, the authors stack multiple unregistered, short-exposure frames from real observations.

While this stacking broadens the Point Spread Function (PSF) due to residual drift, it significantly suppresses stochastic noise, creating a high-SNR baseline that preserves large-scale background structures and instrument signatures.

B. Physics-Based Noise Synthesis Model

The core contribution is a mathematical model (Equation 1) that simulates the full CCD noise formation process. The model synthesizes noise by adding specific components to the high-SNR base:

Signal-Dependent Noise:
- Photon Shot Noise: Modeled as a Poisson distribution based on the expected photoelectron yield, system gain ( $G$ ), and quantum efficiency.
- Photo-Response Non-Uniformity (PRNU): Unlike consumer cameras where PRNU is minor, it is critical in astronomy. The model separates PRNU into:
  - Fixed Pattern: Intrinsic sensor gain variations and etaloning (interference fringes).
  - Donut Artifacts: Out-of-focus shadows from dust on filters/windows.
  - These are estimated from twilight skyflat frames using high-pass filtering and median aggregation.
- Dark Current: Modeled as a per-pixel rate ( $D_{dark}$ ) scaled by exposure time, with Gaussian residuals to account for sample-to-sample variation.
Signal-Independent Noise:
- Read Noise & Fixed Pattern: Modeled using a combination of:
  - A fixed template ( $B_{fp}$ ) derived from stacked bias frames (capturing column-wise banding).
  - Row-wise bias striping (modeled via Gaussian Mixture Models).
  - Pixel-level fluctuations (modeled via Gaussian Mixture Models).
- Hot Pixels & Cosmic Rays: Modeled as a "salt-and-pepper" impulse noise. Hot pixels are fixed defects, while cosmic rays are transient high-energy events. Both are simulated using a Bernoulli distribution with a uniform amplitude range.

C. Training Pipeline

The synthesized noisy images are paired with the high-SNR stacked bases to train a standard U-Net architecture. The model is trained using an L1 loss function.

3. Key Contributions

Physics-Based Noise Formulation: A comprehensive mathematical model tailored specifically to the noise characteristics of scientific CCDs (including PRNU, column banding, and cosmic rays), distinct from consumer CMOS models.
Synthesis Pipeline: A method to generate abundant, realistic paired training datasets (Noisy RAW $\leftrightarrow$ High-SNR Base) using only a limited number of calibration frames and unregistered exposures.
MuSCAT Observer Dataset: The creation of a real-world, multi-band dataset from two twin ground-based telescopes (MuSCAT-3 and MuSCAT-4), providing raw, processed, and calibration frames for rigorous evaluation.
Scientific Accuracy Focus: The framework prioritizes photometric fidelity (preserving flux measurements) over simple visual smoothness, which is critical for scientific analysis.

4. Experimental Results

The authors evaluated their method on the MuSCAT-3 (training/in-domain) and MuSCAT-4 (zero-shot/cross-instrument) datasets across G1, R1, and I1 bands.

Quantitative Performance:
- Photometric Fidelity: The proposed method achieved the highest PSNR on synthetic data and the highest FLUX-SNR (Signal-to-Noise Ratio of injected mock PSFs) on real data, outperforming both traditional methods (BM3D, Drizzle) and state-of-the-art learning-based baselines (N2N, PMN, PNNP).
- Noise Suppression: It achieved lower NMAD (Normalized Median Absolute Deviation) and STD in background regions compared to learning-based baselines, indicating effective noise removal without oversmoothing.
Zero-Shot Generalization: The model trained on MuSCAT-3 successfully denoised MuSCAT-4 data without retraining, demonstrating robustness across different instruments and observing conditions.
Ablation Studies: Removing specific noise components (PRNU, Read Noise, Shot Noise) from the synthesis model significantly degraded performance, confirming that modeling the full physics of CCD noise is essential for accurate flux recovery.
Visual Quality: The method preserved stellar PSF structures and avoided the "oversmoothing" artifacts common in traditional denoisers, which often attenuate peak intensities.

5. Significance

This work bridges the gap between computational imaging and astronomical science. By providing a physically interpretable noise model and a practical way to generate training data, it enables the application of deep learning to scientific imaging where ground truth is unavailable.

Scientific Impact: The method improves the detection limits for faint objects by effectively removing stochastic noise while preserving the precise flux measurements required for astrophysics (e.g., exoplanet transit photometry, quasar monitoring).
Methodological Impact: It establishes a new paradigm for training denoising networks in scientific domains where data is scarce, moving away from purely data-driven approaches toward physics-informed synthesis.
Reproducibility: The authors released the code and the MuSCAT dataset, facilitating further research in astronomical image processing.