Optimal mitigation of random telegraph noise for improved photometry at high frame rates
Christopher Layden, Daniel-Rolf Harbeck, Tejus Deo-Dixit, Nathan Lourie, Gabor Furesz, Kevin Burdge · 2026
Read Optimal mitigation of random telegraph noise for improved photometry at high frame rates by Christopher Layden, Daniel-Rolf Harbeck, Tejus Deo-Dixit, Nathan Lourie, Gabor Furesz, Kevin Burdge (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-2605-19022.txt.
Introduction
Christopher Layden and five colleagues cut random telegraph noise in CMOS sky sensors. They test three sensors, the Sony IMX455, Gpixel GSENSE400 and Fairchild HWK4123. On the IMX455 at high gain, telegraph noise rules loud pixels and lifts the floor by over one fifth. Masking loud pixels and a new jump fix each lift faint star signal by over five percent. The fix runs ahead when shots are coarse or loud pixels sit near a star.
Catalog record
| CREATOR | Christopher Layden, Daniel-Rolf Harbeck, Tejus Deo-Dixit, Nathan Lourie, Gabor Furesz, Kevin Burdge |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 67594 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.19022
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Optimal mitigation of random telegraph noise for improved photometry at high frame rates Christopher Layden, Daniel-Rolf Harbeck, Tejus Deo-Dixit, Nathan Lourie, Gabor Furesz, Kevin Burdge Abstract Random telegraph noise (RTN) is a major contributor to read noise in many CMOS image sensors considered for astronomical use. While scientific CMOS image sensors deliver lower read noise than traditional charge-coupled devices, mitigating RTN would widen this gap and enable more precise photometry when using the fast readout rates achievable by CMOS image sensors. We report the levels of RTN in three CMOS image sensors used in astronomical instruments: the Sony IMX455, Gpixel GSENSE400, and Fairchild Imaging HWK4123. For the IMX455 in a high gain mode, RTN is the dominant source of pixels with high read noise and increases the overall read noise floor by ≳ 20 % \gtrsim 20\% . RTN is present in the GSENSE400 and HWK4123 but to smaller effects. We compare two strategies for RTN mitigation: masking pixels exhibiting RTN or using a new algorithm for correcting RTN jumps. For faint ( < 3 e − <3\,\mathrm{e}^{-} /pix/frame) observations of a stellar field with the IMX455, both masking and our algorithm improved the signal-to-noise ratio (SNR) of light curves by > 5 % >5\% on average. Larger improvements were achieved for sources falling on multiple RTN pixels. Our algorithm outperforms masking, especially when the point spread function is undersampled, masked pixels are near the source center, or read noise and shot noise are comparable. In such cases, masking may even deteriorate photometric precision. In other cases, masking remains an effective RTN mitigation technique. We have made available our software for identifying RTN pixels, parametrizing their bias level distributions, and applying our correction algorithm. I Introduction Many recently commissioned or proposed instruments for ultraviolet (UV), optical, and infrared (IR) astronomy employ complementary metal-oxide-semiconductor (CMOS) image sensors. Custom UV-sensitive CMOS sensors have been developed for the upcoming Ultraviolet Transient Astronomy Satellite (ULTRASAT) and Ultraviolet Explorer (UVEX) missions ( 4 ; 14 ) . Commercially available CMOS image sensors have increasingly been selected for ground-based and space-based optical instruments, including proto-Lightspeed on the Magellan Clay Telescope ( 21 ) , the Large Array Survey Telescope ( 28 , LAST;) , the small telescope fleet of Las Cumbres Observatory ( 8 ; 16 , LCO;) , and the Lazuli Space Observatory ( 33 ) . Hybridized CMOS image sensors (notably the HxRG series outlined in 6 ) have become near-ubiquitous for infrared astronomy, finding use in some of today’s premier space telescopes, such as the James Webb Space Telescope ( 31 , JWST;) and the Roman Space Telescope ( 27 ) . Random telegraph noise (RTN; sometimes called “salt-and-pepper” or popcorn noise) has been observed in all of these types of CMOS image sensors and, specifically, in the detectors of JWST, LAST, the LCO small telescope fleet, and the Lazuli Space Observatory ( 14 ; 29 ; 5 ) . If this noise component were removed, then the read noise floor of these instruments could be significantly improved, allowing higher signal-to-noise ratios (SNR) when observing faint scenes or using high frame rates. RTN mitigation is especially salient for space telescopes, as radiation damage is known to increase the prevalence of RTN ( 18 ; 24 ; 2 ) . RTN arises in solid-state transistors from the capture or release of electrons by carrier traps near the Si/SiO 2 interface ( 37 ) . In a charge-coupled device (CCD), the source follower transistor of the output node(s)—where photogenerated charge is converted to a readable voltage—may contain many such traps. The superposition of these traps changing occupancy can result in 1/f noise ( 36 ) , which can dominate the read noise of CCDs ( 17 ) . As device fabrication has improved, allowing for ever-smaller transistors with ever-fewer defects, the number of interface traps in each transistor has fallen. This is true in particular for CMOS image sensors, in which every pixel has its own source follower transistor, typically with size of \lesssim 1\,\mathrm{\mu m} . In scientific CMOS image sensors, each source follower transistor may have zero to a few traps. When only O(1) traps are present, the bias level of a source follower transistor will bounce between a set of discrete values corresponding to the possible occupation states ( 19 ) . In most imaging applications, the signal in each pixel is sampled twice: once before photogenerated charge has been transferred to the gate of the source follower transistor and once after charge transfer. The first reading is subtracted from the second to recover the signal level. In this process, called correlated double sampling (CDS), much excess read noise is removed, and only changes in trap occupation that occur between the samples will affect the signal value ( 37 ) . However, such changes in occupation between samples can still contribute significantly to an image sensor’s read noise, as RTN. The expected bias level distribution of a pixel with one RTN-generating trap is a sum of three Gaussians, as depicted in Fig. 1 a. The central Gaussian, with relative height A , represents bias level readings where trap occupation did not change between samples; its standard deviation \sigma_{r} represents the read noise the pixel would have if the trap was not present. The Gaussians to the left and right represent measurements where the trap has released or captured a carrier, respectively. These Gaussians have relative heights B_{1} and B_{2}=1-A-B_{1} , are equidistant (with spacing d ) from the central Gaussian, and also have standard deviation \sigma_{r} (release or capture could conceivably affect the standard deviation, but this slight effect is not often observed). Depending on the timescales of trap transition and pixel readout, B_{1} can be near zero 37 . The parameters governing the bias level distribution ( A,B_{1},\sigma_{r},d and others if more than one trap is present) can be different for all of the pixels with RTN in a given sensor, depending on the location of the trap in each transistor and the local distribution of dopant atoms ( 23 ; 19 ) . RTN is a significant noise contributor in instances where the spacing d is larger than both \sigma_{r} and the Poissonian fluctuations in the level of photogenerated charge ( \sim\sqrt{\lambda} for a mean photoelectron number \lambda ). Figure 1: a) Histogram of bias levels reported by a simulated image sensor pixel with random telegraph noise. As is typical in image sensors to avoid clipping negative values, a mean bias level of \mu_{b}=100\, ADU is added. The gain of the simulated pixel is set to 10 ADU/e - . The inherent read noise of the pixel \sigma_{r} is broadened by the RTN to a larger value \sigma_{r,tot} . b) Read noise measured in 10,000 pixels of an IMX455 sensor. The long tail, largely consisting of pixels with RTN, significantly increases the RMS read noise. 29 and 5 have developed methods for identifying RTN in pixels of CMOS and hybrid CMOS image sensors. Across such sensors, RTN was identified in \approx 1–5% of pixels. However, the total contribution that RTN makes to the noise budget of CMOS image sensors has not been reported. Because d is often much larger