🔭 astrophysics

Optimal mitigation of random telegraph noise for improved photometry at high frame rates

This paper evaluates random telegraph noise (RTN) levels in three astronomical CMOS sensors and demonstrates that a new algorithm for correcting RTN jumps outperforms traditional pixel masking, particularly for undersampled sources or when noise levels are comparable, thereby enabling more precise high-frame-rate photometry.

Original authors: Christopher Layden, Daniel-Rolf Harbeck, Tejus Deo-Dixit, Nathan Lourie, Gabor Furesz, Kevin Burdge

Published 2026-05-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Christopher Layden, Daniel-Rolf Harbeck, Tejus Deo-Dixit, Nathan Lourie, Gabor Furesz, Kevin Burdge

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

The Problem: The "Popcorn" Noise in Digital Eyes

Imagine you are trying to listen to a very quiet whisper in a room. Suddenly, someone starts popping popcorn in the corner. Pop. Pop. Pop. Sometimes the pops are loud; sometimes they are quiet. You can't tell if a sound is the whisper you are looking for or just a piece of popcorn hitting the floor.

In the world of astronomy, cameras (specifically CMOS sensors) are the "ears" listening to the faint light of distant stars. However, these cameras suffer from a glitch called Random Telegraph Noise (RTN). The authors call this "popcorn noise" or "salt-and-pepper noise."

Inside the tiny transistors of the camera, electrons get stuck or released by tiny traps, causing the pixel's reading to suddenly jump up or down. This isn't a real star getting brighter or dimmer; it's just the camera glitching. This noise makes it hard to see faint objects, especially when the camera is reading images very quickly (high frame rates).

The Investigation: Testing Three Cameras

The researchers tested three different types of camera sensors used in telescopes:

Sony IMX455: A high-resolution sensor used in many modern telescopes.
Gpixel GSENSE400: A sensor with larger pixels, often used as an upgrade for older cameras.
Fairchild Imaging HWK4123: A super-sensitive sensor designed to count individual photons (particles of light).

They found that all three had this "popcorn" noise, but the Sony IMX455 had the worst of it. In this camera, the noise was so bad that it increased the overall "static" (read noise) of the image by more than 20%. The other two cameras had the noise, but it was much quieter.

The Old Solution vs. The New Solution

When astronomers noticed a pixel was glitching, they used to just mask it.

The Masking Analogy: Imagine you are looking at a painting, but there is a smudge on the glass in front of one specific spot. The old way was to put a black sticker over that spot so you couldn't see the smudge. You lose that part of the picture, but the rest looks clean.
The Problem: If the smudge is right in the middle of a star you are trying to study, putting a black sticker over it ruins your measurement. You lose data.

The authors developed a new algorithm (a smart computer program) to fix this.

The New Solution Analogy: Instead of putting a sticker over the smudge, the program acts like a detective. It looks at the history of that specific pixel. It knows, "This pixel usually sits at level 100, but sometimes it jumps to 110 or 90." When it sees a jump, it says, "Ah, that's just popcorn noise!" and gently nudges the number back to where it should be. It fixes the glitch without erasing the data.

What They Found

The researchers tested this new "detective" program on real images of stars and even a rare event where an asteroid blocked a star's light (an occultation).

Better Clarity: For faint stars, the new program improved the clarity of the data by more than 5% on average. That might sound small, but in astronomy, it's like turning a blurry photo into a sharp one.
When Masking Fails: The old "sticker" method (masking) worked okay in some cases, but when a glitchy pixel was right in the center of a star, masking actually made the data worse. The new program fixed the glitch without losing the star's light.
The Occultation Test: In the asteroid test, the glitchy pixels accidentally made the asteroid look like it blocked more light than it actually did. The new program corrected this, giving a much more accurate measurement of the asteroid's size.

The Takeaway

The authors have made their software free and open to everyone. It allows astronomers to:

Identify which pixels in their cameras are "glitchy."
Fix the glitches in real-time or after the fact.
Get clearer, more precise measurements of the universe without throwing away any data.

In short, they taught the cameras how to ignore their own "popcorn" noise, allowing us to hear the quietest whispers from the stars more clearly.

Technical Summary: Optimal Mitigation of Random Telegraph Noise for Improved Photometry at High Frame Rates

Problem Statement
Random Telegraph Noise (RTN), also known as "salt-and-pepper" or popcorn noise, is a significant source of read noise in Complementary Metal-Oxide-Semiconductor (CMOS) image sensors used for astronomy. While scientific CMOS (sCMOS) sensors generally offer lower read noise than traditional Charge-Coupled Devices (CCDs), the presence of RTN degrades this advantage, particularly in high-frame-rate and low-flux applications. RTN arises from the capture and release of electrons by carrier traps near the Si/SiO2 interface in source follower transistors. In pixels with RTN, the bias level fluctuates between discrete states, creating a long tail in the read noise distribution. This increases the Root Mean Square (RMS) read noise significantly, even if the median read noise remains low, thereby reducing the Signal-to-Noise Ratio (SNR) for faint sources. Existing mitigation strategies, such as masking defective pixels or applying simple median filters, often fail to account for the specific bias level distributions of individual pixels, can degrade photometric precision when the Point Spread Function (PSF) is undersampled, or may introduce systematic biases.

Methodology
The authors investigated three commercially available CMOS sensors used in astronomical instruments: the Sony IMX455 (in QHY600M cameras), the Gpixel GSENSE400 (in QHY42 cameras), and the Fairchild Imaging HWK4123 (in Hamamatsu ORCA-Quest 2 cameras).

Characterization: The team collected thousands of bias frames to measure read noise and identify RTN pixels. They utilized an Anderson-Darling normality test to detect non-Gaussian bias distributions. For pixels failing this test, they fitted a triple-Gaussian model to characterize the bias level distribution, extracting parameters for the central Gaussian (mean $\mu_b$ , standard deviation $\sigma_r$ ) and the side peaks representing RTN jumps (amplitudes $B_1, B_2$ , and jump spacing $d$ ).
Correction Algorithm: A new Python-based algorithm was developed to identify and correct RTN jumps on a per-pixel, per-frame basis.
- Reference Generation: The algorithm estimates the expected signal level for a pixel using either a temporal median (for time-series data) or a spatial median (for single frames), adjusted for the pixel's specific bias parameters.
- Jump Identification: Using a Poisson-Gaussian cumulative density function (CDF), the algorithm calculates significance thresholds ( $\alpha = 0.003$ ) to determine if a pixel's signal deviates significantly from the expected central Gaussian, accounting for the inherent right-skew of Poisson statistics.
- Correction: If a jump is identified, the signal is shifted by the jump magnitude $d$ (converted to electrons). The algorithm includes a bias correction step to prevent the artificial inflation of average flux values caused by the asymmetric correction of positive vs. negative jumps.
Validation: The authors compared their algorithm against a simple "masking" strategy (treating RTN pixels as defective) using Monte Carlo simulations and real astronomical observations. Real-world testing included stellar field photometry and an asteroid occultation event observed with the Las Cumbres Observatory (LCO) small telescope network.

Key Contributions

Quantification of RTN Impact: The study quantifies the contribution of RTN to the total read noise floor. For the Sony IMX455, RTN increases the RMS read noise by $\gtrsim 20\%$ . For the GSENSE400 and HWK4123, the increases are smaller ( $\gtrsim 4\%$ and $\gtrsim 2.5\%$ , respectively), though still relevant for ultra-low noise applications.
Algorithm Development: The authors present a robust, per-pixel correction algorithm that adapts to the specific bias level distribution of each RTN pixel and the local photon flux, avoiding the pitfalls of fixed-threshold methods.
Comparative Analysis: The paper provides a detailed comparison between the proposed correction algorithm and pixel masking, demonstrating that the algorithm outperforms masking, particularly when the PSF is undersampled, masked pixels are near the source center, or read noise is comparable to shot noise.
Open-Source Software: The authors released the RTN-Corrector software suite, including scripts for identifying RTN pixels (rtn_fitter) and applying corrections (rtn_fixer), making these techniques accessible to the astronomical community.

Results

Simulation: Monte Carlo simulations showed that the correction algorithm consistently improved SNR compared to masking. Masking was found to degrade SNR in cases where RTN pixels fell near the core of a source, especially for undersampled PSFs.
Stellar Photometry: In observations of a stellar field using the IMX455, both the correction algorithm and masking improved the SNR of light curves by an average of $>5\%$ for faint sources ( $< 3 e^-$ /pix/frame). The correction algorithm showed lower variance in performance and maintained absolute photometric accuracy, whereas masking tended to underestimate brightness.
Asteroid Occultation: During an occultation of the Trojan asteroid Anchises, the correction algorithm reduced the RMS residuals of the light curve fit by $\approx 7\%$ and corrected a systematic overestimation of the occultation depth caused by an imbalance of negative RTN jumps during the event.
Sensor Specifics: The study noted that the IMX455 exhibited symmetric RTN distributions (both positive and negative jumps), while the GSENSE400 and HWK4123 showed highly asymmetric distributions (mostly positive jumps) and smaller jump magnitudes. The HWK4123's ultra-low capacitance design resulted in RTN jumps smaller than a single photoelectron in many cases, limiting the efficacy of post-processing correction for that specific sensor.

Significance and Claims
The paper claims that mitigating RTN is essential for realizing the full potential of sCMOS sensors, particularly for high-speed photometry and observations of faint objects where read noise dominates. The authors assert that their algorithm provides a reliable method to recover photometric precision that would otherwise be lost to RTN, offering a superior alternative to simple pixel masking. They emphasize that while masking is effective in specific regimes (well-sampled PSFs, read-noise dominated), the correction algorithm offers more consistent improvements across a wider range of conditions, including undersampled PSFs and varying source brightness. The work is presented as a practical tool for astronomers using detectors with RTN, offering an immediate boost in data quality without requiring hardware changes. The authors also note that RTN mitigation is particularly salient for space telescopes, where radiation damage is known to increase the prevalence of RTN.