Pith. sign in

REVIEW 3 major objections 4 minor 25 references

Recognition and classification of the cosmic-ray events in images captured by CMOS/CCD cameras

T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Ordinary camera images carry enough shape information to classify cosmic-ray hits as muon-like, electron-like, or noise, and the muon-like track orientation estimates azimuth.

desk verdict A transparent preliminary methods note from CREDO; the azimuth-estimation claim is the weak spot and needs validation before it becomes a capability. read the letter →

arxiv 1909.01929 v2 pith:JGT3NE7R submitted 2019-09-04 astro-ph.IM

Michal Niedzwiecki , Krzysztof Rzecki , Marta Marek , Piotr Homola , Katarzyna Smelcerz , David Alvarez Castillo , Karel Smolek , Bohdan Hnatyk
show 10 more authors
Jilberto Zamora-Saa Alona Mozgova Vahab Nazaric Dariusz Gora Konrad Kopanski Tadeusz Wibig Alan R. Duffy Jaroslaw Stasielak Zoltan Zimboras Marcin Kasztelan
This is my paper · ORCID
classification astro-ph.IM
keywords cosmic-raydetectionCMOSsensorsCCDmuontracksimagemorphologyellipticitysolidityparticleclassification
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that a single bright trace left in a digital photo can identify a cosmic-ray particle and roughly indicate where it came from. The authors propose that shape features of a hit, especially ellipticity and solidity, separate muon-like tracks from beta-like worm traces and from noise, and that the orientation angle of a muon-like track estimates the muon's azimuth. As proof of concept they process telescope dark frames and a citizen-contributed set of over two million camera images. If the method holds, ordinary smartphones, webcams, and IoT cameras become a worldwide, zero-hardware cosmic-ray observatory.

What carries the argument

The load-bearing object is the per-hit feature vector: ellipticity, marked area, convex area, orientation, track length, and track width, derived from the cropped image and combined with acquisition metadata. Solidity, defined as marked area over convex area, is the companion of ellipticity: the pair separates clean straight tracks from bent worm-like traces and blobs. The same orientation angle used to describe the ellipse bounding the hit becomes, once a hit is classified as muon-like, the estimate of the muon's azimuth. The paper also exploits the fact that hot pixels stay at fixed coordinates from frame to frame, providing a cheap way to cut them out before classification.

What would settle it

A controlled calibration: expose the same camera to a beam of known muons and to a beta source, and check whether muons at shallow angles always give track-like traces (ellipticity above 0.6, solidity above 0.7) and betas always give worm-like traces; any overlap in morphology between the two sources would break the classification. For the azimuth claim, compare track orientation angles against a coincidence telescope co-mounted with the camera; the spread of the residuals measures how much of the orientation is real muon azimuth versus noise.

Watch

Extended reading notes

Core claim

The paper claims that a cosmic-ray hit on a CMOS/CCD sensor can be recognized and typed from the geometry of the trace it leaves. Spot-like hits (ellipticity near 1 and solidity near 1), track-like hits (ellipticity above 0.6 and solidity above 0.7) attributed to muons crossing the sensor plane at an angle, and worm-like hits (ellipticity around 0.5 and solidity below 0.7) attributed to beta radiation are the three morphological kinds. A machine-learning classifier fed with these shape features plus metadata such as camera model, pixel size, RAW versus down-sampled format, GPS location, timestamp, and exposure assigns each cropped hit to muon-like, electron-like, or other; for muon-like events the orientation angle of the track is read as the muon's azimuth. The authors present this as a concept and preliminary implementation, tested on telescope dark frames and on a citizen-contributed set of more than two million camera images.

Load-bearing premise

The load-bearing premise is that trace shape maps one-to-one to particle identity—a high-ellipticity, high-solidity track means a muon, a worm-like trace means beta radiation, and a spot could be either—and that the trace's orientation angle equals the muon's azimuth, a mapping the paper asserts without calibrating it against known particle sources or simulations.

Editorial extensions

If this is right

  • If the shape-to-particle mapping holds, a camera's ordinary images become cosmic-ray science data with no new hardware, only software.
  • Muon azimuths estimated from single images could feed global anisotropy studies once compass and GPS metadata are sufficiently accurate.
  • Astronomical dark frames, routinely taken for calibration, become a free and large dataset of cosmic-ray interactions.
  • Per-device auto-calibration becomes essential but tractable, since each sensor's pixel size and format change the scale of the same morphological signal.
  • The same pipeline can act as a filter, separating true ionizing-particle hits from hot pixels and fake covered-camera artifacts.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension the paper does not run: calibrate the three morphological classes against a known source, converting 'muon-like' into 'muon'.
  • Since the paper notes that transmission artifacts cluster at exactly 0°, 45°, and 90°, an orientation histogram could double as a per-device artifact veto and as a cross-check on any azimuth anisotropy.
  • If the morphological mapping holds, the same features may carry energy information, so normalizing marked area and track width by pixel size could turn the distributed camera network into a crude energy-resolving detector.
  • The azimuth claim combines two sub-claims: track angle equals muon azimuth on the sensor, and the sensor's absolute orientation is known; with compass errors around 30°, absolute azimuth may need a post-hoc orientation fix from recognizable sky objects in the image.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper presents a concept and preliminary implementation for recognizing cosmic-ray events in CMOS/CCD images, using data from the CREDO project and ESO dark frames. The proposed pipeline detects bright hits, crops them, extracts geometric features (ellipticity, solidity, orientation, length, width) and metadata, and classifies events as muon-like, electron-like, or other/noise using hand-set morphology rules. For muon-like events, the abstract also claims the method estimates the azimuth of the muon track. The paper describes the processing chain, refers to publicly available code, and illustrates the output with histograms from a single CREDO device. The authors explicitly frame the work as a proof-of-concept that needs further validation, particularly in the statistical-analysis section.

Significance. If the central claims could be validated, this would offer an appealing capability: a distributed network of ordinary cameras acting as cosmic-ray muon detectors with directional information. The paper has the virtue of reproducibility: references [15] and [16] provide working code for feature extraction and analysis, and the data sources (CREDO, ESO) are public. However, the capability claims are currently supported only by morphological assumptions, not by quantitative evidence. No classification accuracy, ground truth, confusion matrix, detection efficiency, or azimuth residual is reported, and the paper itself acknowledges that confirmation with radiation sources or simulations is required. The contribution is therefore best seen as an openly documented pipeline whose central claims require demonstration.

major comments (3)
  1. [§3.4 and Abstract] The classification into muon-like, electron-like, and other events rests on untested physical assumptions. The text states 'It suppose that the tack-like hits is caused by muons hits in angle to sensor plane, worm-like is caused by β-radiation,' but no calibration against known particle sources, simulations, or a labeled dataset is provided. The paper reports no confusion matrix, classification efficiency, or false-positive rate. Because the abstract's central claim about identifying muon-like events depends entirely on this morphology-to-particle mapping, the classification step needs at least a basic validation or an explicit rephrasing as an unverified hypothesis.
  2. [§3.6 and Abstract] The claim that the method 'estimates azimuth of a muon track' is not supported by the evidence presented. The bottom-left histogram in Fig. 4 plots track orientations, but there is no comparison to independently known azimuths, no compass/GPS validation, and no azimuth residual or angular-error analysis. The authors themselves note that compass accuracy is often about 30°, that artifacts appear at exactly 0°, 45°, and 90°, and that 'more experiments with physical radiation source or cosmic-ray peak simulations or statistical analysis from more devices are required in order to confirm it.' With the currently shown data, the trace orientation is measured, but the physical azimuth is not demonstrated to be estimated.
  3. [Fig. 4] The statistical analysis is qualitative: the histograms illustrate distributions but do not quantify separation power, selection thresholds, or event counts. In particular, the ellipticity and solidity values that define spot-like, track-like, and worm-like events are given only as illustrative examples in Fig. 3c, not as explicit threshold values or as a reproducible decision rule. This makes it impossible for a reader to reproduce the classification or assess its sensitivity and efficiency.
minor comments (4)
  1. [§3.4] There is a typo: 'tack-like' should be 'track-like'. Also, in §3.6 'mouns' should be 'muons'.
  2. [References] Reference [10] is listed as 'The scikit-image project home page', but the URL points to http://photutils.readthedocs.io/; please make the citation consistent with the actual library used for feature extraction.
  3. [Abstract and §4] The phrase '2 millions images' and the compound 'cameralike' in the abstract are awkward; please rephrase as '2 million images' and 'camera types'.
  4. [Fig. 4] Each panel of Fig. 4 should state the number of events plotted, the device model, and the threshold criteria used to select the subsamples, so that the histograms can be interpreted and reproduced.

Circularity Check

2 steps flagged · score 6.0 of 10

The 'muon-like' class and the claimed azimuth estimate reduce by construction to the image-morphology features used to define them.

  1. self definitional [Section 3.4 and Fig. 3 caption; abstract]
    "It suppose that the tack-like hits is caused by muons hits in angle to sensor plane, worm-like is caused by β-radiation. ... (c) track-like hits: ellipticity > 0.6, solidity > 0.7"

    The class 'muon-like event' is defined purely by morphology: track-like means ellipticity > 0.6 and solidity > 0.7. The physical identification with muons is then asserted by the phrase 'It suppose that the tack-like hits is caused by muons' rather than by any external calibration. Consequently, every statistical statement about 'muon-like' events is, by construction, a statement about blobs passing hand-set shape thresholds; the label and the evidence for the label are the same features.

  2. self definitional [Section 3.6 and abstract]
    "For muon-like events our method estimates azimuth of a muon track. ... Orientation of track will be useful, when correct azimuth is placed in hit metadata. ... Not every smartphone supports compass or its accuracy is ~30 ◦."

    The only azimuth information produced by the method is the image-plane 'orientation' feature extracted from the track-like blob. Since 'muon-like' was itself defined as a track-like blob, the claim that the method estimates the azimuth of a muon track reduces to measuring the orientation of the same shape feature that defined the class. No independent ground truth links that image orientation to a physical muon direction; indeed the paper concedes that correct azimuth requires metadata whose compass accuracy is often only about 30 degrees. The capability claim therefore is a restatement of feature extraction rather than an independently validated measurement.

full rationale

The paper is honestly framed as a concept and preliminary implementation, and much of it is self-contained image analysis: detection, cropping, feature extraction, and simple morphology thresholds. There is no fitted model that is trained on target labels and then used to rederive those labels, and the self-citations to the authors' GitHub scripts [15,16] are implementation references, not load-bearing evidence for the physics claims. The circularity is narrower and specific: the central abstract claim that the method 'estimates azimuth of a muon track' rests on equating 'muon-like' with track-like image morphology and then treating the extracted orientation as a muon azimuth. Sec. 3.4 states the equation by supposition, and Sec. 3.6 acknowledges that confirmation requires radiation-source studies, simulations, or multi-device statistics. Because the prediction is by construction just the orientation feature of the shape class that defines 'muon-like', the azimuth capability claim partially reduces to its own input features. Classification into spot, track, and worm shapes is legitimate image analysis; the circular step is only the unvalidated renaming of those shapes as particle species and the subsequent presentation of the orientation of the renamed class as a physical estimate. For this reason the score is 6: one central prediction reduces by construction, while the broader image-processing pipeline retains independent content.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The pipeline's burden rests on hand-set detection thresholds, hand-set morphology class boundaries, and unvalidated physical assumptions about trace shape meaning and metadata reliability. None of these parameters is fitted with error bars, and no external ground truth is used, so the central claim absorbs all of these assumptions.

free parameters (3)
  • brightness detection threshold = 3x average noise, clamped to 80-160
    Hand-set threshold in the CREDO detector app (Sec 3.2) determines which pixels are cropped as candidate events and therefore enters every later classification.
  • crop window size = default 60 x 60 px
    Hand-set crop radius in the detection app (Sec 3.2) defines the image patch from which features are extracted.
  • class boundary thresholds = ellipticity ~0.5 and ~0.6-0.8; solidity ~0.7
    Hand-chosen morphology cutoffs in Sec 3.4 and 3.6 define spot, worm, track, and muon-like classes; no optimization or error analysis is given.
assumptions (4)
  • domain assumption Track-like image traces are caused by muons hitting the sensor at an angle; worm-like traces are caused by beta radiation; spot-like traces may be caused by muons or other radiation.
    Stated in Sec 3.4 and used to label the three classes; motivated by CORSIKA but not calibrated against known particle types on these sensors.
  • domain assumption Hot pixels appear at the same XY coordinates across a sequence of frames, so they can be identified and removed by comparing frames.
    Sec 3.5 relies on this to exclude hot pixels; standard for astronomical CCDs, but not validated for heterogeneous phone cameras and app processing.
  • domain assumption Compass and GPS metadata are accurate enough to define the azimuth reference for a muon track.
    Sec 3.6 acknowledges compass accuracy is about 30 degrees and GPS is often unavailable indoors; the azimuth claim depends on this metadata.
  • domain assumption Non-standard event frequency, very wide tracks, and unusual colors are reliable evidence of artifact images from badly covered cameras.
    Sec 3.5 lists these heuristics for banning devices, but gives no quantified thresholds or validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Recognition and classification of the cosmic-ray events in images captured by CMOS/CCD cameras." pith.science (2026). https://pith.science/paper/JGT3NE7R

@misc{pith2026190901929,
  author       = {Pith},
  title        = {Pith review of: Recognition and classification of the cosmic-ray events in images captured by CMOS/CCD cameras},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JGT3NE7R}},
  note         = {Machine review of arXiv:1909.01929}
}
read the original abstract

Muons and other ionizing radiation produced by cosmic rays and radiative decays affect CMOS/CCD sensor. When particles colliding with sensors atoms cause specific kind of noise on images recorded by cameras. We present a concept and preliminary implementation of method for recognizing those events and algorithms for image processing and their classification by machine learning. Our method consists of analyzing the shape of traces present in images recorded by a camera sensor and metadata related to an image like camera model, GPS location of camera, vertical and horizontal orientation of a camera sensor, timestamp of image acquisition, and other events recognized near-by sensors. The so created feature vectors are classified as either a muon-like event, an electron-like event or the other event, possibly noise. For muon-like events our method estimates azimuth of a muon track. Source of the data is database of CREDO (Cosmic-Ray Extremely Distributed Observatory) project and ESO (European Southern Observatory) archives. The telescope dark frames from ESO are analysed. CREDO project collected so far over 2 millions images of events from many kinds of cameralike: smartphones camera, laptop webcams and Internet of Things cameras localised around the globe.

Figures

Figures reproduced from arXiv: 1909.01929 by the authors.

Figure 1
Figure 1. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. (a) sharp cosmic ray from dark frame from ESO archive (~12µm LW), (b) more fuzzy cosmic ray on Sony ICX285 sensor (6.45µm LW), c) smartphone RGB camera (~1-2 µm LW), non-RAW image, ~4x down-sampled image, (d) Fuzzy cosmic-ray from ESO, non-linear normalization, (e) non-continues muon track (source: DECO experiment). Cosmic ray track depend on pixel area and pixel depth sensitive (Figure 1d). In science CCD/CMOS sens… view at source ↗
Figure 3
Figure 3. (a) spot-like hits: ellipticity near 1, solidity near 1, (b) worm-like hits: ellipticity ~0.5, solidity < 0.7, (c) track-like hits: ellipticity > 0.6, solidity > 0.7, (d) deblending and join track from the same hit (left), join two hits and mark third hit as separated (center), separated hits (interesting case, it is unknown whether is one hit or two separeted) (right). We can achieve the same effect by simple ellip… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Statistical analysis. 3.6 Statistical analysis Some statistical analysis was presented on [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 24 canonical work pages

  1. [15]

    Niedzwiecki, M., ESO archive cosmix-ray extraction script, accessed 16 July 2019, [https://github.com/credo-science/credo-hackhaton-nkg/]

    work page 2019

  2. [16]

    Niedzwiecki, M., scripts for anaysis images from CREDO project database, accessed 16 July 2019, [https://github.com/dzwiedziu-nkg/credo-analysis/]

    work page 2019

  3. [1]

    Groom, D., Cosmic rays and other nonsense in astronomical CCD imagers, Experimental Astronomy (2002) 14: 45 [doi:10.1023/A:1026196806990]

    work page doi:10.1023/a:1026196806990 2002

  4. [2]

    Measurement of camera image sensor depletion thickness with cosmic rays

    J. Vandenbroucke et. al., Measurement of camera image sensor depletion thickness with cosmic rays, Instrumentation and Detectors 2015 [arXiv:1511.00660]

    work page Pith review arXiv 2015

  5. [3]

    The CREDO main website, accessed 16 July 2019, [https://credo.science/credo-detector-mobile-app/]

    work page 2019

  6. [4]

    ESO Science Archive Facility, accessed 16 July 2019, [ http://archive.eso.org/]

    work page 2019

  7. [5]

    CREDO Detector for Android, accessed 16 July 2019, [https://credo.science/credo-detector-mobile-app/]

    work page 2019

  8. [6]

    CREDO Detector for Windows, accessed 16 July 2019, [https://github.com/credo-science/CREDO-PC-Windows/]

    work page 2019

Show all 25 references
  1. [7]

    CREDO Detector for RaspberryPi, accessed 16 July 2019, [https://github.com/credo-science/ Credo-detector-for-linux-desktop-and-Raspberry-Pi/]

    2019

  2. [8]

    NASA Mars Curiosity rover saw a weird light, but don’t freak out, CNET, accessed 16 July 2019, [https://www.cnet.com/news/ nasa-mars-curiosity-rover-saw-a-weird-light-but-dont-freak-out/] 6 Recognition and classification of the cosmic-ray events in images captured by CMOS/CCD c...

    2019

  3. [9]

    The AstroPy project home page, accessed 16 July 2019, [ https://www.astropy.org/]

    2019

  4. [10]

    The scikit-image project home page, accessed 16 July 2019, [https://photutils.readthedocs.io/]

    2019

  5. [11]

    The scikit-image project home page, accessed 16 July 2019, [ https://scikit-image.org/]

    2019

  6. [12]

    CMS Collaboration, Measurement of the charge ratio of atmospheric muons with the CMS detector, Phys.Lett.B692:83-104,2010 [arXiv:1005.5332]

    2010 arXiv

  7. [13]

    X.; Pulido, J.; Thorman, P.; Hamann, B., An improved method for object detection in astronomical images, MNRAS 451 (4):4445-4459

    Zheng, C. X.; Pulido, J.; Thorman, P.; Hamann, B., An improved method for object detection in astronomical images, MNRAS 451 (4):4445-4459. 2015. [doi:10.1093/mnras/stv1237]

    2015 doi

  8. [14]

    Fisher-Levine and A

    M. Fisher-Levine and A. Nomerotski, Characterising CCDs with cosmic rays, JINST (2015) [doi:10.1088/1748-0221/10/08/c08006]

    2015 doi

  9. [17]

    ACC Cyfronet AGH-UST, API tools for CREOD project, accessed 16 July 2019, [https://github.com/credo-science/credo-api-tools/]

    2019

  10. [18]

    Matthew Meehan et al., The particle detector in your pocket: The Distributed Electronic Cosmic-ray Observatory, Instrumentation and Methods for Astrophysics ICRC 2017 [arXiv:1708.01281]

    2017 arXiv

  11. [19]

    The CRAYFISH main website, accessed 16 July 2019, [ https://crayfis.io/]

    2019

  12. [20]

    Heck and T

    D. Heck and T. Pierog, Extensive Air Shower Simulation with CORSIKA: A Userâ ˘A ´Zs Guide, KIT - Universitat des Landes Baden-W ´l urttemberg und nationales Forschungszentrum in der Helmholtz-Gemeinschaft, accessed 16 July 2019, [https://web.ikp.kit.edu/corsika/usersguide/use...

    2019

  13. [21]

    Huotari, K., Hamari, J., Defining Gamification â ˘A¸ S A Service Marketing Perspective, Proceedings of the 16th International Academic MindTrek Conference 2012Tampere, Finland, October 3-5 [doi:10.1145/2393132.2393137]

  14. [22]

    M. Niedzwieck et.al., Asynchronous Agent System For Monitoring Communication And System States Based On The SOA Paradigm, 25th Conference on Modelling and Simulation, 2011 [ISBN: 978-0-9564944-2-9]

    2011

  15. [23]

    The CosmicWatch project main website, accessed 16 July 2019, [http://www.cosmicwatch.lns.mit.edu/]

    2019

  16. [24]

    Sungsoo Choi, KyungHo Lee, et. al., An all pixel PDAF CMOS image sensor with 0.64Îijmx1.28Îijm photodiode separated by self-aligned in-pixel deep trench isolation for high AF performance Symposium on VLSI Technology 2017 [doi:10.23919/VLSIT.2017.7998212]

    2017

  17. [25]

    C. R. Zhao et al., Adaptive Pixel Crosstalk Compensation for CMOS Image Sensor Advanced Materials Research V ol. 981, pp. 310-314, 2014 [doi:10.4028/www.scientific.net/AMR.981.310] 7

    2014 doi

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.