Pith. sign in

REVIEW 2 major objections 4 minor 60 references

A deep network turns blurry all-sky WISE images into Spitzer-resolution views, halving aperture flux errors and deblending close pairs.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A neural network trained on paired WISE/Spitzer images can sharpen WISE infrared cutouts to Spitzer-like resolution, improving aperture photometry errors by roughly 2x and deblending recovery by roughly 4x over interpolation on held-out COSMOS data.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection The WISE-to-Spitzer idea is worth taking seriously and the paper is careful about many things, but the headline test metrics are likely inflated by train/test spatial overlap in COSMOS, so the generalization claim is not yet established. the 2 major comments →

arxiv 2607.14295 v1 pith:YPPP5WPO submitted 2026-07-15 astro-ph.IM

Enhancing WISE Infrared Imaging to Spitzer Resolution Using Deep Learning Super-Resolution

Saeed Rezaee , Shoubaneh Hemmati , Bahram Mobasher , Naveen A. Reddy , Tara Fetherolf , Alexander de la Vega , Sina Miri This is my paper
classification astro-ph.IM
keywords super-resolutiondeep learningWISESpitzer IRACinfrared imagingaperture photometrysource deblendingCOSMOS
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 attempts to show that a convolutional network trained on roughly 390,000 matched WISE/Spitzer cutouts can genuinely recover Spitzer-level detail from WISE images, which cover the whole sky but have about 3.6 times coarser resolution. The authors report that the trained network measures fixed-aperture integrated flux to a median 11% error, about half the error of interpolation, and recovers about 3.8 times more source peaks at the 3-5 arcsecond separations where WISE blends neighbors that Spitzer can separate. If correct, this would mean all-sky infrared survey data could be sharpened to the resolution of a pointed telescope that observed only selected fields. The paper also characterizes where the model fails, notably oversmoothing of source profiles that biases faint-source fluxes upward.

Core claim

The central claim is that the trained model performs genuine resolution enhancement and deblending rather than interpolation artifacts. From a 14x14 WISE W1 cutout it produces a 64x64 image on the Spitzer pixel grid, and the resulting aperture photometry matches Spitzer truth to a median 11% relative error on integrated flux versus 22% for bicubic interpolation; at 3-5 arcsecond separations it recovers 35% of Spitzer-detectable peaks versus 9% for bicubic. The error grows monotonically toward fainter sources, and the dominant failure mode is oversmoothing that redistributes flux into broad wings.

What carries the argument

The carrier is the Enhanced RCAN, a convolutional network with residual groups and channel attention, a multiscale input block, and progressive subpixel upsampling that maps 2.75-arcsecond WISE pixels to 0.6-arcsecond Spitzer pixels. It is trained with a source-focused composite loss that upweights bright pixels regardless of position, which matters because most cutouts in the dense COSMOS field contain several bright sources; the network learns to reconstruct all of them rather than only the central object.

Load-bearing premise

The held-out test set is treated as a measure of generalization, but the random shuffled split of one small field does not guarantee that the same sky regions do not appear in both training and test; if spatial overlap exists, the reported gains may reflect memorization of specific patches rather than general super-resolution.

What would settle it

Re-evaluate the model with a split that enforces a minimum angular separation (e.g., more than an arcminute) between training and test cutouts, or test on an independent field with Spitzer or JWST imaging not used in training. If aperture flux errors rise above roughly 20% or 3-5 arcsecond peak recovery falls toward the bicubic baseline, the generalization claim is refuted.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • WISE data outside the COSMOS field can be processed to Spitzer-like resolution, since WISE covered the entire sky; the paper shows qualitatively well-behaved output on sky positions without Spitzer coverage.
  • Aperture photometry on enhanced images is about twice as accurate as interpolation, with the largest gain on faint sources (integrated flux error 13% vs 41% for bicubic in the faintest quartile).
  • Source deblending improves by about 3.8x at 3-5 arcsecond separations, turning blended WISE sources into individually measurable peaks.
  • Faint-companion recovery improves by about 1.3x over interpolation, although it stays well below the Spitzer truth ceiling of 63.5%.
  • The oversmoothing failure mode biases integrated fluxes upward, and the paper identifies explicit source-profile, perceptual, or adversarial losses as the natural next mitigations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the model's accuracy holds outside COSMOS, WISE's all-sky coverage could be reprocessed into a uniform Spitzer-resolution infrared survey, enabling time-domain and morphological studies that were previously impossible with WISE alone.
  • Because the train/test split is random within one small field and does not enforce angular separation, the same sky patches could appear in both training and test; a spatially disjoint split would tell whether the reported 11% error and 35% recovery represent generalization or memorization.
  • The learning curve had not flattened at 390k samples, so training on additional Spitzer-covered fields should further reduce errors, and the framework may extend to other WISE/IRAC band pairs with separate validation.
  • The oversmoothing bias could be calibrated by comparing model PSFs to Spitzer PSFs in dense fields, allowing users to apply a flux correction when measuring aperture photometry on enhanced images.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper presents an Enhanced RCAN super-resolution network that maps 14×14 WISE W1 cutouts to 64×64 Spitzer IRAC Ch1 cutouts (4.6× upsampling) using ~390,000 paired cutouts from COSMOS. The model is evaluated on 83,592 held-out cutouts with metrics including SSIM/PSNR, aperture photometry, brightness-binned errors, deblending recall versus source separation, and faint companion recovery against the COSMOS2020 catalog. The central claims are that the model recovers central-source aperture integrated flux to 11% median relative error (vs ~22% for interpolation), recovers 35% of Spitzer-truth peaks at 3–5 arcsec separations (vs ~9% for bicubic), and that performance improves monotonically with source brightness. An appendix replicates the analysis at W2→Ch2 with consistent results. The trained model and code are publicly available.

Significance. If the reported generalization results are valid, this is a practically valuable contribution: WISE covers the entire sky while Spitzer did not, and the claimed factor-of-two photometric improvement and ~4× deblending improvement at 3–5 arcsec separations would enable new science with existing all-sky infrared data. The paper is unusually thorough for a methods paper: it includes multiple baselines, brightness-resolved analysis, a direct deblending measurement, a catalog-based companion test, an end-to-end example, and a second-wavelength replication. The public release of the trained model and code is a real strength. No formal proofs are claimed; the contribution is empirical, so the validity of the evaluation set is the central load-bearing point.

major comments (2)
  1. [Section 3.4 / 2.3] The held-out test set is not spatially independent of the training set. The 70/15/15 split is stratified only by Spitzer peak brightness, and 38.5″ cutouts in the dense COSMOS field mean that the same sky region appears in many cutouts; a random split therefore places overlapping patches on both sides of the boundary. Because the loss upweights bright pixels position-agnostically (Section 3.3), the network is explicitly trained to reconstruct the same local source configurations on which it is then tested. Consequently, the test-set metrics in Tables 2, 5, 6, and 7 may partly measure memorization of specific sky patches rather than generalization to new sky. The out-of-mosaic check (Section 5.7) is only qualitative and has no ground truth, so it does not close this gap. A spatial split (e.g., leaving out a contiguous region) or a quantitative external-field validation is required to supp
  2. [Abstract and §4.3/Table 3] The claim of a monotonic brightness dependence is contradicted by the paper's own numbers. Table 3 lists aperture integrated flux errors of 13.1%, 10.8%, 11.2%, and 8.3% for the faint-to-bright quartiles, and peak errors of 36.6%, 15.7%, 19.9%, and 12.9%. The med-high bin is not between med-low and bright for either metric. The abstract's 'monotonic' statement and the corresponding text should be corrected, or the analysis revised to explain the non-monotonicity and its implications.
minor comments (4)
  1. [Section 3.1.2] The asinh normalization parameters (xsoft, P1, P99) appear to be computed per cutout and stored for inversion. Since the normalization is per-image, a source of fixed physical brightness will be scaled differently depending on the other sources in the cutout. Please clarify whether the normalization is per-cutout or global, and discuss the implications for the learned mapping.
  2. [Section 3.3.1] The term 'position-agnostic' is used to describe the loss weighting, but the weight is a function of pixel intensity (w=3.0 for y>τ, 0.5 otherwise). This is position-independent only in a spatial sense; it is brightness-dependent. Consider renaming to 'position-independent' to avoid confusion.
  3. [Abstract and Table 6] The abstract states '9% for interpolation' in the deblending comparison, but Table 6 reports 9.2% for bicubic and 7.6% for bilinear. Please specify which baseline is being quoted, or average the two with an explicit definition.
  4. [Section 5.8] The Limitations subsection does not mention the potential spatial overlap between training and test cutouts. Once the split is made spatially disjoint (or an external validation is added), this limitation should be explicitly discussed.

Circularity Check

1 steps flagged

Spatial overlap between training and test cutouts in COSMOS makes the 'held-out' generalization metrics partially circular; a spatial split or external-field test is needed.

specific steps
  1. fitted input called prediction [Section 2.3 (cutout extraction), Section 3.4 (70/15/15 split); metrics in Tables 2, 5, 6, 7]
    "We split the dataset 70/15/15 (stratified by Spitzer truth peak brightness) into a training pool of 390,091, a validation set of 83,592 used to select the best model during training, and a held-out test set of 83,592 used exclusively for the metrics reported in Section 4."

    The split is stratified only by Spitzer peak brightness, with no spatial separation. Cutouts are 38.5″ across and ~557,000 are drawn uniformly from the 2 deg² COSMOS field, so the aggregate cutout footprint is ~32× the field area; a typical sky patch appears on both sides of the split. The position-agnostic loss (Sec. 3.3.1) explicitly 'trains to reconstruct all bright pixels in the field rather than focusing on the central catalog primary,' so the central source of a test cutout is also a bright neighbor in many training cutouts. The 'held-out' predictions can therefore be produced by recalling sky patches already seen in training; Tables 2, 5, 6, and 7 measure overlap/memorization, not generalization to genuinely new sky. Sec. 5.7's out-of-mosaic check is explicitly only qualitative ('a

full rationale

The core supervised-learning pipeline is not circular in itself: the network is trained on paired WISE/Spitzer images and tested on cutouts it did not train on as cutouts. There is no self-definitional equation, no fitted parameter renamed as a prediction in the usual sense, and no load-bearing self-citation. The circularity is in the evaluation design: because COSMOS sources and 38.5″ cutouts are split randomly with stratification only by peak brightness and no spatial separation, the same sky area appears in both training and test cutouts. Given the source density and ~390k training cutouts, the reported test metrics can be substantially inflated by memorization of specific sources, so the paper's central claim—that the model performs genuine 4.6× super-resolution that will transfer to new sky—is not established by Tables 2, 5, 6, and 7. The paper itself acknowledges that the only external check (Sec. 5.7) is qualitative and that quantitative validation outside COSMOS 'is left to future work' (Sec. 5.8). A spatial split (e.g., disjoint sky regions for train/val/test) or a quantitative test on external Spitzer/JWST imaging would remove this circularity. Because the network and code are public and the baselines are legitimate, the flaw is fixable and the score reflects partial, not total, circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central claim rests on standard supervised-learning assumptions plus a small set of hand-set hyperparameters. No new physical entities are introduced. The most consequential assumption is that the COSMOS-based training set and the held-out split support generalization; the spatial-leakage issue makes that assumption fragile.

free parameters (6)
  • source threshold tau = 0.5
    Pixels above tau in normalized Spitzer truth get weight 3.0 in the loss; chosen by hand in Section 3.3.1. It shapes how the model treats faint versus bright pixels.
  • loss component weights (alpha, beta, gamma) = 0.5, 0.35, 0.15
    Hand-set in Section 3.3.4; the authors state they did not perform an exhaustive search. They rank pixel fidelity, SSIM, and gradient terms.
  • progressive supervision weights (lambda_28, lambda_56, lambda_64) = 0.2, 0.3, 0.5
    Hand-set in Section 3.3.4; candidates were compared on validation but no exhaustive search.
  • deblending peak detection threshold = 1.0 in normalized space
    Used in Section 5.5 for the deblending experiment; the authors note the qualitative finding is robust to threshold variation, but the absolute recall numbers depend on it.
  • companion detection peak threshold = 0.3 in normalized space
    Used in Section 5.6 for faint companion detection; a methodological choice that affects recall levels.
  • asinh softening parameter xsoft = median of positive pixel values
    Defines the normalization stretch in Section 3.1.2; computed from each image's data but is a chosen statistic, not derived from a theory.
axioms (4)
  • domain assumption WISE W1 and Spitzer IRAC Ch1 bandpasses are close enough that color terms are <0.1 mag for most galaxy SEDs
    Invoked in Section 2.2 to neglect color corrections. If false, the learned mapping includes a color-dependent bias.
  • domain assumption The WISE->Spitzer mapping is learnable from paired examples and the unWISE W1 PSF is well-enough behaved for the network to invert
    The entire supervised-learning setup assumes the resolution gap can be bridged with training data; stated implicitly in Section 1 and 3.
  • domain assumption The COSMOS field is representative of the sky regions where the model will be applied
    The paper trains and tests on COSMOS only; the out-of-field test in Section 5.7 is qualitative. Acknowledged in Section 5.8 for dense fields and compact sources.
  • domain assumption Source variability between WISE and Spitzer observations is negligible
    WISE and Spitzer images were taken at different epochs; variable AGN or transient sources would add mismatch noise that the model cannot learn. Not discussed in the paper.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing WISE Infrared Imaging to Spitzer Resolution Using Deep Learning Super-Resolution." pith.science (2026). https://pith.science/paper/YPPP5WPO

@misc{pith2026260714295,
  author       = {Pith},
  title        = {Pith review of: Enhancing WISE Infrared Imaging to Spitzer Resolution Using Deep Learning Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YPPP5WPO}},
  note         = {Machine review of arXiv:2607.14295}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We present a deep-learning framework that performs 4.6x spatial super-resolution from WISE W1 (3.4 micron) toward Spitzer IRAC Ch1 (3.6 micron), and characterize its behavior on the COSMOS field. Our sample consists of ~390,000 paired cutouts drawn uniformly within the WISE/Spitzer overlap, with a held-out test set of 83,592 cutouts on which we report all metrics. The framework uses a convolutional neural network (an Enhanced Residual Channel Attention Network) trained with a loss function that emphasizes accurate recovery of sources in crowded fields. The model recovers the total flux of the central source in a fixed aperture to a median relative error of 11%, a factor of ~2 better than the interpolation baselines; the gain reaches ~3x on the faintest quartile. The brightness dependence is monotonic: the aperture integrated flux error decreases from 13% on the faintest quartile to 8% on the brightest. At the 3-5 arcsec separations where WISE blends sources that Spitzer separates, the model recovers 35% of the source peaks detectable in the Spitzer truth compared with 9% for interpolation. The characteristic failure mode is oversmoothing of source profiles, which biases integrated flux measurements upward; this pattern is qualitatively similar to that of the interpolation baselines but is quantitatively smaller for the trained model. These results suggest genuine resolution enhancement and source deblending, providing a path toward applying super-resolution across the all-sky area that Spitzer could not cover. An appendix replicates the analysis at W2 -> IRAC Ch2 with consistent results; the trained model and code are publicly available.

Figures

Figures reproduced from arXiv: 2607.14295 by Alexander de la Vega, Bahram Mobasher, Naveen A. Reddy, Saeed Rezaee, Shoubaneh Hemmati, Sina Miri, Tara Fetherolf.

Figure 1
Figure 1. Figure 1: Coverage of the unWISE W1 tiles contributing to the paired dataset, with per-tile counts of sampled catalog sources in the legend. The shaded region with the dashed gray outline marks the Spitzer-COSMOS Ch1 mosaic footprint; blue dashed boxes mark the 1.56◦×1.56◦ boundaries of the four contributing unWISE tiles; points show the sampled source positions, colored by the tile that supplied the WISE cutout. A … view at source ↗
Figure 2
Figure 2. Figure 2: Representative paired WISE → Spitzer cutouts from the test set, spanning faint, mid-brightness, bright, and blended regimes (peak intensities annotated; values in normalized units). Top: WISE W1 inputs (14×14 pixels at 2.75′′/pixel). Bottom: Spitzer IRAC Ch1 ground truth (64×64 pixels at 0.6′′/pixel). Each pair covers the same 38.4′′ patch of sky. Each cutout is centered on a COSMOS2020 catalog source; bec… view at source ↗
Figure 3
Figure 3. Figure 3: COSMOS2020 catalog source density across the COSMOS field. The density is sufficient that the majority of 38.5′′ cutouts contain multiple catalog neighbors, motivating the position-agnostic loss design in Section 3.3. Source counts use the quality-filtered catalog subset (IRAC Ch1 SNR > 5, galaxies, MAG < 25), the same selection as the companion analysis in Section 5.6; the training sample itself applies n… view at source ↗
Figure 4
Figure 4. Figure 4: Resolution gap between WISE W1 (3.4 µm, FWHM ≈ 6 ′′) and Spitzer IRAC Ch1 (3.6 µm, FWHM ≈ 1.7 ′′) illustrated on four representative paired cutouts. The solid green circles mark the FWHM of each instrument’s point-spread function, centered on the catalog position (cyan cross). The 3–4× FWHM ratio is what the network must invert [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Architecture of the Enhanced Residual Channel Attention Network (Enhanced RCAN) used for super-resolution from WISE W1 (14 × 14) to Spitzer IRAC Ch1 (64 × 64). The model consists of a multiscale feature extraction block, four residual groups with channel attention (32 RCABs total), and progressive subpixel upsampling with intermediate supervision at 28×28 and 56×56 tap-offs [PITH_FULL_IMAGE:figures/full_f… view at source ↗
Figure 6
Figure 6. Figure 6: The loss upweights pixels that belong to genuine sources relative to the background. (a) Spitzer truth for a representative test cutout in normalized intensity; the cyan line on the color bar marks the source threshold τ . (b) Per-pixel loss weight w(yi): pixels above τ receive wsource = 3.0 (red), all others receive wbackground = 0.5 (gray). The upweighted pixels cover ≈ 27% of this cutout and are distrib… view at source ↗
Figure 7
Figure 7. Figure 7: Representative super-resolution results across the test set. Each row is one WISE → Spitzer test cutout, spanning a resolved galaxy, a blended close pair, a crowded field, and a moderate multi-source field. Columns show the WISE W1 input, the bicubic and Simple CNN baselines, the Enhanced RCAN output, and the Spitzer IRAC Ch1 ground truth. The four 64×64 panels in each row share one display stretch. The En… view at source ↗
Figure 8
Figure 8. Figure 8: Aperture photometry at the central catalog position: predicted (y-axis) versus Spitzer ground truth (x-axis) for each method. Top: peak aperture (r = 2 px). Bottom: integrated aperture (r = 6 px). The dashed line in each panel marks the 1:1 relation. The Enhanced RCAN (right column) hugs the 1:1 line tightly. The vertical scatter differs substantially between methods. We quantify the dispersion with the no… view at source ↗
Figure 9
Figure 9. Figure 9: Metrics binned by brightness across the four methods. Each panel shows one metric, with one bar group per brightness bin and one bar per method. The flux panels show the median relative error in aperture flux at the central catalog position, at r = 2 px for the peak and r = 6 px for the integrated aperture, the same definitions as Tables 3 and 4. The RCAN advantage is largest in the faint and intermediate … view at source ↗
Figure 10
Figure 10. Figure 10: Enhanced RCAN reconstruction of a representative bright source cutout, compared with the baseline methods. Top row: the WISE input, the three super-resolution methods, and the Spitzer ground truth. The Spitzer truth resolves multiple distinct sources across the field, and the Enhanced RCAN output preserves the same sources with comparable sharpness. The bicubic upsampling smooths every source into a broad… view at source ↗
Figure 11
Figure 11. Figure 11: End-to-end application to a single COSMOS2020 galaxy (ID 376723, z = 0.34). Left: the WISE W1 input cutout. Middle: the Enhanced RCAN output. Right: the Spitzer IRAC Ch1 ground truth. The green circle marks the integrated aperture (r = 3.6 ′′) at the catalog position. Fluxes are measured after inverting the normalization to physical units and subtracting the local sky in an annulus. The model recovers the… view at source ↗
Figure 12
Figure 12. Figure 12: Deblending peak recovery recall versus minimum source separation, for the four methods. The Enhanced RCAN consistently recovers ≈ 4× more of the peaks detectable in the Spitzer truth than bicubic upsampling across the deblending regime (2–10′′), indicating that the model performs real source separation rather than interpolating WISE PSF wings [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Faint companion detection recall versus (separation from primary, catalog flux ratio) for each method. The Enhanced RCAN (bottom-left) substantially exceeds the interpolation baselines (top row) across the entire (separation, ratio) grid; the Spitzer truth oracle (bottom-right) sets the upper bound on what is detectable in this setup [PITH_FULL_IMAGE:figures/full_fig_p023_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Comparison of WISE input (left), bicubic upsampling (middle), and Enhanced RCAN output (right) for four sky positions inside unWISE tile coverage but outside the Spitzer-COSMOS mosaic. No Spitzer ground truth is available for any of these positions. Rows are ordered by increasing input intensity. The Enhanced RCAN output is qualitatively consistent with the model’s output inside the COSMOS field, suggesti… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

60 extracted references · 7 canonical work pages

  1. [1]

    2015, TensorFlow : Large-Scale Machine Learning on Heterogeneous Systems, https://www.tensorflow.org/

    Abadi , M., et al. 2015, TensorFlow : Large-Scale Machine Learning on Heterogeneous Systems, https://www.tensorflow.org/

    2015

  2. [2]

    2022, The Astrophysical Journal, 935, 167, 10.3847/1538-4357/ac7c74

    Astropy Collaboration . 2022, The Astrophysical Journal, 935, 167, 10.3847/1538-4357/ac7c74

    doi:10.3847/1538-4357/ac7c74 2022

  3. [3]

    1996, , 117, 393, 10.1051/aas:1996164

    Bertin , E., & Arnouts , S. 1996, , 117, 393, 10.1051/aas:1996164

    doi:10.1051/aas:1996164 1996

  4. [4]

    C., et al

    Calzetti , D., Armus , L., Bohlin , R. C., et al. 2000, , 533, 682, 10.1086/308692

    doi:10.1086/308692 2000

  5. [5]

    M., Narayanan , D., & Cooray , A

    Casey , C. M., Narayanan , D., & Cooray , A. 2014, Physics Reports, 541, 45, 10.1016/j.physrep.2014.02.009

    doi:10.1016/j.physrep.2014.02.009 2014

  6. [6]

    2001, , 556, 562, 10.1086/321609

    Chary , R., & Elbaz , D. 2001, , 556, 562, 10.1086/321609

    doi:10.1086/321609 2001

  7. [7]

    M., Wright , E

    Cutri , R. M., Wright , E. L., Conrow , T., et al. 2013, Explanatory Supplement to the AllWISE Data Release Products , IPAC/Caltech

    2013

  8. [8]

    2022, , 939, L4, 10.3847/2041-8213/ac98af

    Dabbech , A., Terris , M., Jackson , A., et al. 2022, , 939, L4, 10.3847/2041-8213/ac98af

    doi:10.3847/2041-8213/ac98af 2022

  9. [9]

    J., & Asensio Ramos , A

    D \'i az Baso , C. J., & Asensio Ramos , A. 2018, Astronomy & Astrophysics, 614, A5, 10.1051/0004-6361/201731344

    work page doi:10.1051/0004-6361/201731344 2018

  10. [10]

    W., & Dambre , J

    Dieleman , S., Willett , K. W., & Dambre , J. 2015, , 450, 1441, 10.1093/mnras/stv632

    doi:10.1093/mnras/stv632 2015

  11. [11]

    C., He , K., & Tang , X

    Dong , C., Loy , C. C., He , K., & Tang , X. 2014, in Proceedings of the European Conference on Computer Vision (ECCV) , 184--199, 10.1007/978-3-319-10593-2_13

    work page doi:10.1007/978-3-319-10593-2_13 2014

  12. [12]

    T., & Li , A

    Draine , B. T., & Li , A. 2007, , 657, 810, 10.1086/511055

    doi:10.1086/511055 2007

  13. [13]

    S., et al

    Elbaz , D., Dickinson , M., Hwang , H. S., et al. 2011, , 533, A119, 10.1051/0004-6361/201117239

    doi:10.1051/0004-6361/201117239 2011

  14. [14]

    G., Hora , J

    Fazio , G. G., Hora , J. L., Allen , L. E., et al. 2004, The Astrophysical Journal Supplement Series, 154, 10, 10.1086/422843

    doi:10.1086/422843 2004

  15. [15]

    S., & Hook , R

    Fruchter , A. S., & Hook , R. N. 2002, , 114, 144, 10.1086/338393

    doi:10.1086/338393 2002

  16. [16]

    2018, Monthly Notices of the Royal Astronomical Society, 480, 3749, 10.1093/mnras/sty2102

    Gheller , C., & Vazza , F. 2018, Monthly Notices of the Royal Astronomical Society, 480, 3749, 10.1093/mnras/sty2102

    doi:10.1093/mnras/sty2102 2018

  17. [17]

    R., Millman , K

    Harris , C. R., Millman , K. J., van der Walt , S. J., et al. 2020, Nature, 585, 357, 10.1038/s41586-020-2649-2

    doi:10.1038/s41586-020-2649-2 2020

  18. [18]

    Hausen , R., & Robertson , B. E. 2020, , 248, 20, 10.3847/1538-4365/ab8868

    doi:10.3847/1538-4365/ab8868 2020

  19. [19]

    2018, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 7132--7141, 10.1109/CVPR.2018.00745

    Hu , J., Shen , L., Albanie , S., Sun , G., & Wu , E. 2018, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 7132--7141, 10.1109/CVPR.2018.00745

    arXiv 2018

  20. [20]

    2023, , 40, e001, 10.1017/pasa.2022.55

    Huertas-Company , M., & Lanusse , F. 2023, , 40, e001, 10.1017/pasa.2022.55

    doi:10.1017/pasa.2022.55 2023

  21. [21]

    Hunter , J. D. 2007, Computing in Science and Engineering, 9, 90, 10.1109/MCSE.2007.55

    doi:10.1109/mcse.2007.55 2007

  22. [22]

    H., Cohen , M., Masci , F., et al

    Jarrett , T. H., Cohen , M., Masci , F., et al. 2011, The Astrophysical Journal, 735, 112, 10.1088/0004-637X/735/2/112

    doi:10.1088/0004-637x/735/2/112 2011

  23. [23]

    2005, Annual Review of Astronomy and Astrophysics, 43, 727, 10.1146/annurev.astro.43.072103.150606

    Lagache , G., Puget , J.-L., & Dole , H. 2005, Annual Review of Astronomy and Astrophysics, 43, 727, 10.1146/annurev.astro.43.072103.150606

    Pith/arXiv arXiv 2005

  24. [24]

    2014, The Astronomical Journal, 147, 108, 10.1088/0004-6256/147/5/108

    Lang , D. 2014, The Astronomical Journal, 147, 108, 10.1088/0004-6256/147/5/108

    doi:10.1088/0004-6256/147/5/108 2014

  25. [25]

    2021, , 507, 1546, 10.1093/mnras/stab2195

    Lauritsen , L., Dickinson , H., Bromley , J., et al. 2021, , 507, 1546, 10.1093/mnras/stab2195

    work page doi:10.1093/mnras/stab2195 2021

  26. [26]

    M., Mackay , C

    Law , N. M., Mackay , C. D., & Baldwin , J. E. 2006, , 446, 739, 10.1051/0004-6361:20053695

    doi:10.1051/0004-6361:20053695 2006

  27. [27]

    2017, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 105--114, 10.1109/CVPR.2017.19

    Ledig , C., Theis , L., Husz \'a r , F., et al. 2017, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 105--114, 10.1109/CVPR.2017.19

    doi:10.1109/cvpr.2017.19 2017

  28. [28]

    2021, in Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW) , 1833--1844, 10.1109/ICCVW54120.2021.00210

    Liang , J., Cao , J., Sun , G., et al. 2021, in Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW) , 1833--1844, 10.1109/ICCVW54120.2021.00210

    arXiv 2021

  29. [29]

    Lim , B., Son , S., Kim , H., Nah , S., & Lee , K. M. 2017, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , 136--144, 10.1109/CVPRW.2017.151

    doi:10.1109/cvprw.2017.151 2017

  30. [30]

    2019, in International Conference on Learning Representations (ICLR)

    Loshchilov , I., & Hutter , F. 2019, in International Conference on Learning Representations (ICLR)

    2019

  31. [31]

    Lucy , L. B. 1974, , 79, 745, 10.1086/111605

    doi:10.1086/111605 1974

  32. [32]

    H., Gunn , J

    Lupton , R. H., Gunn , J. E., & Szalay , A. S. 1999, The Astronomical Journal, 118, 1406, 10.1086/301004

    doi:10.1086/301004 1999

  33. [33]

    M., et al

    Mainzer , A., Bauer , J., Cutri , R. M., et al. 2014, The Astrophysical Journal, 792, 30, 10.1088/0004-637X/792/1/30

    doi:10.1088/0004-637x/792/1/30 2014

  34. [34]

    M., Lang , D., & Schlegel , D

    Meisner , A. M., Lang , D., & Schlegel , D. J. 2017, The Astronomical Journal, 154, 161, 10.3847/1538-3881/aa894e

    doi:10.3847/1538-3881/aa894e 2017

  35. [35]

    J., Shuntov , M., et al

    Moneti , A., McCracken , H. J., Shuntov , M., et al. 2022, Astronomy & Astrophysics, 658, A126, 10.1051/0004-6361/202142361

    doi:10.1051/0004-6361/202142361 2022

  36. [36]

    1986, , 24, 127, 10.1146/annurev.aa.24.090186.001015

    Narayan , R., & Nityananda , R. 1986, , 24, 127, 10.1146/annurev.aa.24.090186.001015

    arXiv 1986

  37. [37]

    R., & Smirnov , O

    Offringa , A. R., & Smirnov , O. 2017, , 471, 301, 10.1093/mnras/stx1547

    doi:10.1093/mnras/stx1547 2017

  38. [38]

    Y., Ho , L

    Peng , C. Y., Ho , L. C., Impey , C. D., & Rix , H.-W. 2002, , 124, 266, 10.1086/340952

    doi:10.1086/340952 2002

  39. [39]

    P., Massa , P., Kinakh , V., et al

    Ramunno , F. P., Massa , P., Kinakh , V., et al. 2025, , 698, A140, 10.1051/0004-6361/202453205

    work page doi:10.1051/0004-6361/202453205 2025

  40. [40]

    Rau , U., & Cornwell , T. J. 2011, , 532, A71, 10.1051/0004-6361/201117104

    doi:10.1051/0004-6361/201117104 2011

  41. [41]

    W., Parul , H., & Gleyzer , S

    Reddy , P., Toomey , M. W., Parul , H., & Gleyzer , S. 2024, Machine Learning: Science and Technology, 5, 035076, 10.1088/2632-2153/ad76f8

    work page doi:10.1088/2632-2153/ad76f8 2024

  42. [42]

    Richardson , W. H. 1972, Journal of the Optical Society of America, 62, 55, 10.1364/JOSA.62.000055

    doi:10.1364/josa.62.000055 1972

  43. [43]

    B., & Mirabel , I

    Sanders , D. B., & Mirabel , I. F. 1996, , 34, 749, 10.1146/annurev.astro.34.1.749

    doi:10.1146/annurev.astro.34.1.749 1996

  44. [44]

    B., Salvato , M., Aussel , H., et al

    Sanders , D. B., Salvato , M., Aussel , H., et al. 2007, The Astrophysical Journal Supplement Series, 172, 86, 10.1086/517885

    doi:10.1086/517885 2007

  45. [45]

    Schawinski , K., Zhang , C., Zhang , H., Fowler , L., & Santhanam , G. K. 2017, Monthly Notices of the Royal Astronomical Society, 467, L110, 10.1093/mnrasl/slx008

    doi:10.1093/mnrasl/slx008 2017

  46. [46]

    2007, The Astrophysical Journal Supplement Series, 172, 1, 10.1086/516585

    Scoville , N., Aussel , H., Brusa , M., et al. 2007, The Astrophysical Journal Supplement Series, 172, 1, 10.1086/516585

    doi:10.1086/516585 2007

  47. [47]

    2016, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 1874--1883, 10.1109/CVPR.2016.207

    Shi , W., Caballero , J., Husz \'a r , F., et al. 2016, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 1874--1883, 10.1109/CVPR.2016.207

    doi:10.1109/cvpr.2016.207 2016

  48. [48]

    2024, , 686, A272, 10.1051/0004-6361/202349100

    Song , W., Ma , Y., Sun , H., Zhao , X., & Lin , G. 2024, , 686, A272, 10.1051/0004-6361/202349100

    work page doi:10.1051/0004-6361/202349100 2024

  49. [49]

    J., & Navarro , F

    Sukurdeep , Y., Budav \'a ri , T., Connolly , A. J., & Navarro , F. 2025, , 170, 233, 10.3847/1538-3881/adfb72

    work page doi:10.3847/1538-3881/adfb72 2025

  50. [50]

    F., Valtchanov , I., Lieu , M., et al

    Sweere , S. F., Valtchanov , I., Lieu , M., et al. 2022, , 517, 4054, 10.1093/mnras/stac2437

    doi:10.1093/mnras/stac2437 2022

  51. [51]

    2024, , 681, A118, 10.1051/0004-6361/202347048

    Traina , A., Gruppioni , C., Delvecchio , I., et al. 2024, , 681, A118, 10.1051/0004-6361/202347048

    doi:10.1051/0004-6361/202347048 2024

  52. [52]

    L., Nunez-Iglesias , J., et al

    van der Walt , S., Sch \"o nberger , J. L., Nunez-Iglesias , J., et al. 2014, PeerJ, 2, e453, 10.7717/peerj.453

    doi:10.7717/peerj.453 2014

  53. [53]

    2018, in Proceedings of the European Conference on Computer Vision (ECCV) Workshops , 10.1007/978-3-030-11021-5\_5

    Wang , X., Yu , K., Wu , S., et al. 2018, in Proceedings of the European Conference on Computer Vision (ECCV) Workshops , 10.1007/978-3-030-11021-5\_5

    doi:10.1007/978-3-030-11021-5 2018

  54. [54]

    C., Sheikh , H

    Wang , Z., Bovik , A. C., Sheikh , H. R., & Simoncelli , E. P. 2004, IEEE Transactions on Image Processing, 13, 600, 10.1109/TIP.2003.819861

    arXiv 2004

  55. [55]

    R., Kauffmann , O

    Weaver , J. R., Kauffmann , O. B., Ilbert , O., et al. 2022, , 258, 11, 10.3847/1538-4365/ac3078

    doi:10.3847/1538-4365/ac3078 2022

  56. [56]

    W., Roellig , T

    Werner , M. W., Roellig , T. L., Low , F. J., et al. 2004, The Astrophysical Journal Supplement Series, 154, 1, 10.1086/422992

    doi:10.1086/422992 2004

  57. [57]

    L., Eisenhardt , P

    Wright , E. L., Eisenhardt , P. R. M., Mainzer , A. K., et al. 2010, The Astronomical Journal, 140, 1868, 10.1088/0004-6256/140/6/1868

    doi:10.1088/0004-6256/140/6/1868 2010

  58. [58]

    B., & Lahav , O

    Zaroubi , S., Hoffman , Y., Fisher , K. B., & Lahav , O. 1995, , 449, 446, 10.1086/176070

    doi:10.1086/176070 1995

  59. [59]

    2018, in Proceedings of the European Conference on Computer Vision (ECCV) , 294--310, 10.1007/978-3-030-01234-2\_18

    Zhang , Y., Li , K., Li , K., et al. 2018, in Proceedings of the European Conference on Computer Vision (ECCV) , 294--310, 10.1007/978-3-030-01234-2\_18

    doi:10.1007/978-3-030-01234-2 2018

  60. [60]

    2017, IEEE Transactions on Computational Imaging, 3, 47, 10.1109/TCI.2016.2644865

    Zhao , H., Gallo , O., Frosio , I., & Kautz , J. 2017, IEEE Transactions on Computational Imaging, 3, 47, 10.1109/TCI.2016.2644865

    arXiv 2017

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.