pith. sign in

arxiv: 2606.29255 · v1 · pith:6ZPNEAYBnew · submitted 2026-06-28 · 💻 cs.CV · cs.AI

Confidence-feedback-weighted graph matching network: online-offline laser-induced damage site matching under complex interference

Yueyue Han , Guanhua Chen , Hangcheng Dong , Kang Zhang , Fengdong Chen , Zhitao Peng , Fa Zeng , Qihua Zhu
show 1 more author
Guodong Liu
This is my paper

Pith reviewed 2026-06-30 08:07 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords graph matchingdamage site matchinglaser-induced damageconfidence feedbackdistractor suppressionpoint set matchingonline-offline alignmentcomplex scene matching
0
0 comments X

The pith

A graph matching network feeds back estimated node confidence as edge weights to suppress distractors when matching laser damage sites from coordinates alone.

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

The paper presents a graph matching method for aligning online inspection images of laser optics with offline reference images, where many pseudo-damage sites create distractors that closely resemble true damage. It claims that estimating each node's matchability confidence after every round of scoring and then using those values to weight the edges during feature aggregation reduces the influence of incorrect matches and sharpens discrimination across the two graphs. The method adds a geometric consistency check to correct overconfident errors and a hard-example loss to focus training on difficult pairs, all while using only the centroid coordinates of the sites as input. Experiments on a custom dataset of complex scenes report a matching F1-score of 96.36 percent with stable runtime.

Core claim

The central claim is that a confidence-feedback-weighted graph matching network, which estimates node matchability confidence from successive matching scores and feeds those confidences back as reliability weights to guide edge-feature aggregation, suppresses distractor propagation and improves cross-graph discriminability. A geometric consistency constraint calibrates spurious high-confidence estimates, and a hard-example mining loss sharpens distinction among structurally similar sites. The network operates on damage-site centroid coordinates alone and reaches 96.36 percent F1-score on the Complex-Scene dataset.

What carries the argument

The confidence-feedback-weighted graph matching network, which estimates per-node matchability confidence after each scoring round and re-uses it as an edge reliability weight in the next aggregation step.

If this is right

  • Only centroid coordinates are required as input, so the method works even when richer visual features are unavailable or unreliable.
  • The feedback of matchability confidence limits the spread of distractor effects through the graph.
  • The added geometric consistency constraint corrects many overconfident but incorrect matches.
  • Hard-example mining improves separation between sites that share similar local geometry.
  • The approach yields 96.36 percent F1-score while remaining computationally efficient on the reported dataset.

Where Pith is reading between the lines

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

  • The same feedback mechanism could be tested on other point-set matching tasks that suffer from many look-alike distractors, such as tracking features across video frames or aligning sensor readings in robotics.
  • Because the method depends on geometric consistency to correct early mistakes, performance may degrade in scenes where true sites violate the assumed spatial regularity.
  • Iterating the confidence feedback for more rounds than shown in the paper might further reduce residual errors in extremely cluttered images.

Load-bearing premise

That the confidence values derived from early matching scores are accurate enough to guide later steps without locking in and amplifying initial errors caused by distractors.

What would settle it

A controlled test set in which distractors are crafted to receive high initial match scores; if the final F1-score on that set falls well below 96.36 percent, the feedback loop is not suppressing errors as claimed.

Figures

Figures reproduced from arXiv: 2606.29255 by Fa Zeng, Fengdong Chen, Guanhua Chen, Guodong Liu, Hangcheng Dong, Kang Zhang, Qihua Zhu, Yueyue Han, Zhitao Peng.

Figure 1
Figure 1. Figure 1: Schematic illustration of online-offline damage-site matching. Red regions denote true damage sites with corresponding matches, whereas green regions denote detected damage candidates without true counterparts, referred to as non-corresponding sites. This task is considerably more challenging than conven￾tional image matching due to several complex factors, as 1 arXiv:2606.29255v1 [cs.CV] 28 Jun 2026 [PIT… view at source ↗
Figure 2
Figure 2. Figure 2: Typical challenges in online-offline damage-site matching. Red regions indicate matched true damage sites. Green regions indicate non￾corresponding damage sites. Cyan regions denote online damage sites reprojected onto the offline image. illustrated in [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Classical message-passing mechanisms of GNNs. ω21, F(e21), and α21 denote the structural normalization weight in GCN, the edge￾conditioned weight in ECC, and the attention weight in GAT, respectively; e21 is the edge feature from node 2 to node 1 [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Schematic illustration of suppressing interference from non-corresponding sites via the iterative confidence-feedback weighting mechanism in CFW-GMN [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The pipeline of CFW-GMN.The variables in the figure are consistent with those defined in the main text. aggregation. This stage encodes neighborhood geometric relationships, described by relative displacement and orien￾tation information, into node representations and provides initial features for subsequent matching inference. The uniform aggregation weight is defined as: w (1) iu = 1 |Nr (vi)| . (2) Matc… view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of edge-feature aggregation weights. Red regions indicate matched true damage sites. Green regions indicate non￾corresponding damage sites. tions may share similar local structures, leading to feature ambiguity. To enhance the global positional awareness of node representations, this study introduces positional encod￾ing based on random Fourier features[42,43]. Specifically, the random Fourie… view at source ↗
Figure 7
Figure 7. Figure 7: (a) Illustration of the receptive fields of global and local attention mechanisms. (b) Visualization of the attention weights generated by local self-attention and cross-attention. 2.3. Optimal transport-based matching inference In CFW-GMN, this module is also executed three times. Note that only the second and third iterations perform match￾ing inference and output predicted correspondences, whereas the f… view at source ↗
Figure 8
Figure 8. Figure 8: Illustration of ambiguity-prone local matching. The matching scores of distractor sites can be close to, or even higher than, those of the true correspondences, thereby causing local matching ambiguity. Taking the offline branch as an example, Moff denotes the set of ground-truth damaged-site indices with local hard samples. For each i ∈ Moff, nearby interfering sites within radius dhard are collected as N… view at source ↗
Figure 9
Figure 9. Figure 9: Representative samples from the simulated training set and the real-data test set. (a) Spatially Non-uniform Damage Distribution. (b) Interference from Non-corresponding Damage Sites. (c) Morphological Deformation of Damage Sites. (d) Local positional offset [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Performance comparison of different matching methods. (a) Bar chart of the F1-score standard deviation across different samples. (b) Radar chart of multi-metric matching performance [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Matching results of CFW-GMN on the Complex-Scene dataset. Damage sites with the same color at corresponding online-offline locations indicate predicted correspondences, while repeated colors may appear due to the limited color palette [PITH_FULL_IMAGE:figures/full_fig_p010_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Visualization of edge-feature weights during feature aggregation in Stages 2 and 3. Red regions indicate matched true damage sites. Green regions indicate non-corresponding damage sites [PITH_FULL_IMAGE:figures/full_fig_p011_12.png] view at source ↗
read the original abstract

Online inspection images of final optics in high-power laser facilities contain pseudo-damage sites that closely resemble true damage sites. Determining the authenticity of online-detected sites is therefore difficult and requires accurate matching to offline ground-truth sites. However, this matching remains highly challenging due to limited match-discriminative features, local geometric distortions, and numerous distractor sites. Existing matching models mainly suppress distractors implicitly through loss-function supervision. We propose a confidence-feedback-weighted graph matching network that requires only damage-site centroid coordinates as input. It estimates node matchability confidence from each round of matching scores and feeds it back as a reliability weight to guide subsequent edge-feature aggregation, thereby suppressing distractor propagation and enhancing cross-graph discriminability. Within this framework, a geometric consistency constraint calibrates spurious high-confidence matchability estimates, while a hard-example mining loss improves discrimination between structurally similar sites. Experiments on our Complex-Scene dataset show that the proposed method achieves a matching F1-score of 96.36$\%$ with robust and efficient performance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes a confidence-feedback-weighted graph matching network for online-offline matching of laser-induced damage sites. It takes only centroid coordinates as input, iteratively estimates node matchability confidence from matching scores and feeds these back as reliability weights to guide edge-feature aggregation, applies a geometric consistency constraint to calibrate spurious estimates, and uses a hard-example mining loss. On the authors' Complex-Scene dataset the method reports a 96.36% F1-score.

Significance. If the feedback mechanism can be shown to suppress distractor propagation without amplifying early errors, the approach would offer a lightweight, coordinate-only solution to a practical inspection problem in high-power laser facilities. The geometric-consistency and hard-mining components are standard, so the novelty and performance gain rest entirely on the stability of the confidence-feedback loop.

major comments (2)
  1. [Method] Method section (description of iterative feedback): the claim that the geometric consistency constraint 'calibrates spurious high-confidence matchability estimates' is not accompanied by an explicit formulation or timing diagram showing whether the constraint is enforced inside the same iteration before the feedback weight is applied; without this, the risk that an initial false high-confidence distractor pair is reinforced remains unaddressed.
  2. [Experiments] Experiments section: no ablation isolating the confidence-feedback component versus a non-iterative baseline is reported, nor are dataset statistics (number of distractors per scene, coordinate noise levels) or error bars on the 96.36% F1 provided, making it impossible to verify that the reported gain is attributable to the feedback mechanism rather than the hard-example mining loss alone.
minor comments (1)
  1. [Abstract] The abstract states 'requires only damage-site centroid coordinates as input' yet later refers to 'edge-feature aggregation'; clarify whether any additional features beyond centroids are ever used.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and will incorporate clarifications and additional experiments in the revised version.

read point-by-point responses

  1. Referee: [Method] Method section (description of iterative feedback): the claim that the geometric consistency constraint 'calibrates spurious high-confidence matchability estimates' is not accompanied by an explicit formulation or timing diagram showing whether the constraint is enforced inside the same iteration before the feedback weight is applied; without this, the risk that an initial false high-confidence distractor pair is reinforced remains unaddressed.

    Authors: We agree that the original description is insufficiently precise on the iteration timing. In the revision we will add an explicit mathematical formulation of the geometric consistency constraint (including the calibration step) together with a timing diagram that shows the constraint being enforced inside each iteration before the matchability confidence is fed back as an edge weight. This will make clear that spurious estimates are calibrated prior to influencing subsequent aggregation steps. revision: yes

  2. Referee: [Experiments] Experiments section: no ablation isolating the confidence-feedback component versus a non-iterative baseline is reported, nor are dataset statistics (number of distractors per scene, coordinate noise levels) or error bars on the 96.36% F1 provided, making it impossible to verify that the reported gain is attributable to the feedback mechanism rather than the hard-example mining loss alone.

    Authors: We acknowledge the absence of these controls. For the revised manuscript we will add (1) an ablation that compares the full iterative confidence-feedback model against a non-iterative baseline that uses only the hard-example mining loss, (2) dataset statistics reporting the average number of distractors per scene and the coordinate noise levels present in the Complex-Scene dataset, and (3) error bars (mean ± standard deviation over multiple random seeds) for the 96.36% F1-score. These additions will allow readers to isolate the contribution of the feedback loop. revision: yes

Circularity Check

0 steps flagged

No circularity: new architecture proposal with no self-referential derivations

full rationale

The paper introduces a novel graph matching network architecture that uses node matchability confidence feedback and a geometric consistency constraint. No equations, derivations, or parameter-fitting steps are shown that reduce the claimed F1-score or matching performance to inputs by construction. The method is presented as an empirical proposal relying on standard GNN components plus feedback weighting, without any self-definitional loops, fitted-input predictions, or load-bearing self-citations that collapse the central claim. The derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; the method relies on standard graph neural network assumptions and a new dataset whose construction details are absent.

pith-pipeline@v0.9.1-grok · 5737 in / 1074 out tokens · 30518 ms · 2026-06-30T08:07:00.895632+00:00 · methodology

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

47 extracted references · 38 canonical work pages

  1. [1]

    Target Breakthrough Enabled Fusion Record at NIF

    C. Osolin, “Target Breakthrough Enabled Fusion Record at NIF”, https://lasers.llnl.gov/news/target- breakthrough-enabled-fusion-record-nif (April 15, 2026)

    2026

  2. [2]

    Ultrafast Process Characterization of Laser-Induced Damage in Fused Silica Using Pump- Probe Shadow Imaging Techniques

    Z. Liu, J. Zhang, S. Wang, F. Geng, Q. Zhang, J. Cheng, M. Chen, and Q. Xu, “Ultrafast Process Characterization of Laser-Induced Damage in Fused Silica Using Pump- Probe Shadow Imaging Techniques”, Materials 17, 837 (2024). DOI: https://doi.org/10.3390/ma17040837

    work page doi:10.3390/ma17040837 2024

  3. [3]

    Full-scale optic designed for onsite study of damage growth at the Laser MegaJoule facility

    C. Lacombe, L. Lamaign `ere, G. Hallo, M. Sozet, T. Donval, G. Raz’e, C. Ameil, M. Benoit, F. Gaudfrin, E. Bordenave, N. Bonod, and J. N’eauport, “Full-scale optic designed for onsite study of damage growth at the Laser MegaJoule facility”, Opt. Express 31, 4291–4305 (2023). DOI: https://doi.org/10.1364/OE.474581

    work page doi:10.1364/oe.474581 2023

  4. [4]

    Analysis of optics damage growth at the National Ignition Facility

    Z. M. Liao, M. Nostrand, P. Whitman, and J. Bude, “Analysis of optics damage growth at the National Ignition Facility”, Proc. SPIE 9632, 963217 (2015). DOI: https://doi.org/10.1117/12.2195515

    work page doi:10.1117/12.2195515 2015

  5. [5]

    The Prediction of Incremental 12 Y.Han et al. Damage on Optics from the Final Optic Assembly in an ICF High-Power Laser Facility

    X. Hu, W. Zhou, H. Guo, X. Huang, B. Zhao, W. Zhong, Q. Zhu, and Z. Chen, “The Prediction of Incremental 12 Y.Han et al. Damage on Optics from the Final Optic Assembly in an ICF High-Power Laser Facility”, Appl. Sci. 14, 5226 (2024). DOI: https://doi.org/10.3390/app14125226

    work page doi:10.3390/app14125226 2024

  6. [6]

    Final Optics Damage Inspection (FODI) for the National Ignition Facility

    A. D. Conder, J. J. Chang, L. M. Kegelmeyer, M. L. Spaeth, and P. K. Whitman, “Final Optics Damage Inspection (FODI) for the National Ignition Facility”, Proc. SPIE 7797, 77970P (2010). DOI: https://doi.org/10.1117/12.862596

    work page doi:10.1117/12.862596 2010

  7. [7]

    Automated optics inspection analysis for NIF

    L. M. Kegelmeyer, R. Clark, R. R. Leach, D. McGuigan, V . M. Kamm, D. Potter, J. T. Salmon, J. Senecal, A. Conder, M. Nostrand, and P. K. Whitman, “Automated optics inspection analysis for NIF”, Fusion Eng. Des. 87, 2120–2124 (2012). DOI: https://doi.org/10.1016/j.fusengdes.2012.09.017

    work page doi:10.1016/j.fusengdes.2012.09.017 2012

  8. [8]

    Final optics damage online inspection in high power laser facility

    F. P. Wei, F. D. Chen, J. Tang, Z. T. Peng, and G. D. Liu, “Final optics damage online inspection in high power laser facility”, Optoelectron. Lett. 15, 306–311 (2019). DOI: https://doi.org/10.1007/s11801-019-8193-3

    work page doi:10.1007/s11801-019-8193-3 2019

  9. [9]

    Automatic classification of true and false laser-induced damage in large aperture optics

    F. Wei, F. Chen, B. Liu, Z. Peng, J. Tang, Q. Zhu, D. Hu, Y . Xiang, N. Liu, Z. Sun, and G. Liu, “Automatic classification of true and false laser-induced damage in large aperture optics”, Opt. Eng. 57, 053112 (2018). DOI: https://doi.org/10.1117/1.OE.57.5.053112

    work page doi:10.1117/1.oe.57.5.053112 2018

  10. [10]

    Inspection and Repair of Optical Damage in Tradition and Deep Learning (Invited)

    Y . Li, J. L. Li, Z. Li, D. A. Liu, D. W. Zhang, and J. Y . Zhang, “Inspection and Repair of Optical Damage in Tradition and Deep Learning (Invited)”, Acta Photonica Sin. 51, 1012002 (2022). DOI: https://doi.org/10.3788/gzxb20225110.1012002

    work page doi:10.3788/gzxb20225110.1012002 2022

  11. [11]

    Laser performance of the SG-III laser facility

    W. Zheng, X. Wei, Q. Zhu, F. Jing, D. Hu, J. Su, K. Zheng, X. Yuan, H. Zhou, W. Dai, W. Zhou, F. Wang, D. Xu, X. Xie, B. Feng, Z. Peng, L. Guo, Y . Chen, X. Zhang, L. Liu, D. Lin, Z. Dang, Y . Xiang, and X. Deng, “Laser performance of the SG-III laser facility”, High Power Laser Sci. Eng. 4, e21 (2016). DOI: https://doi.org/10.1017/hpl.2016.20

    work page doi:10.1017/hpl.2016.20 2016

  12. [12]

    Injection laser system architecture, upgrades, and future at the National Ignition Fa- cility

    J. E. Heebner, “Injection laser system architecture, upgrades, and future at the National Ignition Fa- cility”, Proc. SPIE 13358, 133580F (2025). DOI: https://doi.org/10.1117/12.3040268

    work page doi:10.1117/12.3040268 2025

  13. [13]

    Fusion enabling laser-induced damage reduction, management, and repair strategies at the National Ignition Facility

    C. W. Carr, “Fusion enabling laser-induced damage reduction, management, and repair strategies at the National Ignition Facility”, Proc. SPIE 12726, 1272602 (2023). DOI: https://doi.org/10.1117/12.2688417

    work page doi:10.1117/12.2688417 2023

  14. [14]

    Laser-induced damage growth of large- aperture fused silica optics under high-fluence 351 nm laser irradiation

    X. Chai, P. Li, J. Zhao, G. Wang, D. Zhu, Y . Jiang, B. Chen, Q. Zhu, B. Feng, L. Wang, and Y . Jing, “Laser-induced damage growth of large- aperture fused silica optics under high-fluence 351 nm laser irradiation”, Optik 226, 165549 (2021). DOI: https://doi.org/10.1016/j.ijleo.2020.165549

    work page doi:10.1016/j.ijleo.2020.165549 2021

  15. [15]

    Surface weak scratch detection for optical elements based on a multimodal imaging system and a deep encoder–decoder network

    X. Liang, J. Sun, X. Wang, J. Li, L. Zhang, and J. Guo, “Surface weak scratch detection for optical elements based on a multimodal imaging system and a deep encoder–decoder network”, J. Opt. Soc. Am. A 40, 1237–1248 (2023). DOI: https://doi.org/10.1364/JOSAA.483381

    work page doi:10.1364/josaa.483381 2023

  16. [16]

    Optics Recycle Loop Strategy for NIF Operations Above UV Laser-Induced Damage Threshold

    M. L. Spaeth, P. J. Wegner, T. I. Suratwala, M. C. Nostrand, J. D. Bude, A. D. Conder, J. A. Folta, J. E. Heebner, L. M. Kegelmeyer, B. J. MacGowan, D. C. Mason, M. J. Matthews, and P. K. Whitman, “Optics Recycle Loop Strategy for NIF Operations Above UV Laser-Induced Damage Threshold”, Fusion Sci. Technol. 69, 265–294 (2016). DOI: https://doi.org/10.1318...

    work page doi:10.13182/fst15-119 2016

  17. [17]

    Distinctive image features from scale-invariant keypoints

    D. G. Lowe, “Distinctive image features from scale-invariant keypoints”, Int. J. Comput. Vis. 60, 91–110 (2004). DOI: https://doi.org/10.1023/B:VISI.0000029664.99615.94

    work page doi:10.1023/b:visi.0000029664.99615.94 2004

  18. [18]

    SURF: Speeded Up Robust Features

    H. Bay, T. Tuytelaars, and L. Van Gool, “SURF: Speeded Up Robust Features”, inComputer Vision – ECCV 2006, Lecture Notes Comput. Sci. 3951, 404–417 (2006). DOI: https://doi.org/10.1007/11744023 32

    work page doi:10.1007/11744023 2006

  19. [19]

    Faster and bet- ter: A machine learning approach to corner detection

    E. Rosten, R. Porter, and T. Drummond, “Faster and bet- ter: A machine learning approach to corner detection”, IEEE Trans. Pattern Anal. Mach. Intell. 32, 105–119 (2010). DOI: https://doi.org/10.1109/TPAMI.2008.275

    work page doi:10.1109/tpami.2008.275 2010

  20. [20]

    BRIEF: Computing a Local Binary Descriptor Very Fast

    M. Calonder, V . Lepetit, M. ¨Ozuysal, T. Trzcinski, C. Strecha, and P. Fua, “BRIEF: Computing a Local Binary Descriptor Very Fast”, IEEE Trans. Pattern Anal. Mach. Intell. 34, 1281–1298 (2012). DOI: https://doi.org/10.1109/TPAMI.2011.222

    work page doi:10.1109/tpami.2011.222 2012

  21. [22]

    URLhttps://doi.org/10.1109/ICCV.2019.00943

    S. Leutenegger, M. Chli, and R. Y . Siegwart, “BRISK: Binary Robust Invariant Scalable Keypoints”, inProc. IEEE Int. Conf. Comput. Vis.2548–2555 (2011). DOI: https://doi.org/10.1109/ICCV .2011.6126542

    work page doi:10.1109/iccv 2011

  22. [23]

    Fast Explicit Diffusion for Accelerated Features in Nonlinear Scale Spaces

    P. F. Alcantarilla, J. Nuevo, and A. Bartoli, “Fast Explicit Diffusion for Accelerated Features in Nonlinear Scale Spaces”, inProc. Br. Mach. Vis. Conf.13.1–13.11 (2013). DOI: https://doi.org/10.5244/C.27.13

    work page doi:10.5244/c.27.13 2013

  23. [24]

    A Survey of Lost-in-Space Star Identification Algorithms Since 2009

    D. Rijlaarsdam, H. Yous, J. Byrne, D. Oddenino, G. Furano, and D. Moloney, “A Survey of Lost-in-Space Star Identification Algorithms Since 2009”, Sensors 20, 2579 (2020). DOI: https://doi.org/10.3390/s20092579

    work page doi:10.3390/s20092579 2009

  24. [25]

    A Survey on Star Identification Algorithms

    B. B. Spratling IV and D. Mortari, “A Survey on Star Identification Algorithms”, Algorithms 2, 93–107 (2009). DOI: https://doi.org/10.3390/a2010093

    work page doi:10.3390/a2010093 2009

  25. [26]

    The Pyramid Star Identification Technique

    D. Mortari, M. A. Samaan, C. Bruccoleri, and J. L. Junkins, “The Pyramid Star Identification Technique”, Navigation 51, 171–183 (2004). DOI: https://doi.org/10.1002/j.2161-4296.2004.tb00349.x

    work page doi:10.1002/j.2161-4296.2004.tb00349.x 2004

  26. [27]

    Geometric voting algorithm for star trackers

    M. Kolomenkin, S. Pollak, I. Shimshoni, and M. Lindenbaum, “Geometric voting algorithm for star trackers”, IEEE Trans. Aerosp. Electron. Syst. 44, 441–456 (2008). DOI: https://doi.org/10.1109/TAES.2008.4560198 CFW-GMN 13

    work page doi:10.1109/taes.2008.4560198 2008

  27. [28]

    Star pattern identification technique by modified grid algorithm

    H. Lee and H. Bang, “Star pattern identification technique by modified grid algorithm”, IEEE Trans. Aerosp. Electron. Syst. 43, 1112–1116 (2007). DOI: https://doi.org/10.1109/TAES.2007.4383600

    work page doi:10.1109/taes.2007.4383600 2007

  28. [29]

    Full-sky autonomous star identification based on radial and cyclic features of star pattern

    G. Zhang, X. Wei, and J. Jiang, “Full-sky autonomous star identification based on radial and cyclic features of star pattern”, Image Vis. Comput. 26, 891–897 (2008). DOI: https://doi.org/10.1016/j.imavis.2007.10.006

    work page doi:10.1016/j.imavis.2007.10.006 2008

  29. [30]

    Spherical Polar Pattern Matching for Star Identification

    J. Fu, L. Lin, and Q. Li, “Spherical Polar Pattern Matching for Star Identification”, Sensors 25, 4201 (2025). DOI: https://doi.org/10.3390/s25134201

    work page doi:10.3390/s25134201 2025

  30. [31]

    The Graph Neural Network Model

    F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfardini, “The Graph Neural Network Model”, IEEE Trans. Neural Netw. 20, 61–80 (2009). DOI: https://doi.org/10.1109/TNN.2008.2005605

    work page doi:10.1109/tnn.2008.2005605 2009

  31. [32]

    Graph Matching Networks for Learning the Similarity of Graph Structured Objects

    Y . Li, C. Gu, T. Dullien, O. Vinyals, and P. Kohli, “Graph Matching Networks for Learning the Similarity of Graph Structured Objects”, inProc. 36th Int. Conf. Mach. Learn., Proc. Mach. Learn. Res. 97, 3835–3845 (2019). https://proceedings.mlr.press/v97/li19d.html

    2019

  32. [33]

    Neural Message Passing for Quantum Chemistry

    J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl, “Neural Message Passing for Quantum Chemistry”, inProc. 34th Int. Conf. Mach. Learn., Proc. Mach. Learn. Res. 70, 1263–1272 (2017). https://proceedings.mlr.press/v70/gilmer17a.html

    2017

  33. [34]

    Semi-Supervised Classification with Graph Convolutional Networks

    T. N. Kipf and M. Welling, “Semi-Supervised Classification with Graph Convolutional Networks”, inInt. Conf. Learn. Represent.(2017). https://openreview.net/forum?id=SJU4ayYgl

    2017

  34. [35]

    Dynamic Edge- Conditioned Filters in Convolutional Neural Net- works on Graphs

    M. Simonovsky and N. Komodakis, “Dynamic Edge- Conditioned Filters in Convolutional Neural Net- works on Graphs”, inProc. IEEE Conf. Com- put. Vis. Pattern Recognit.3693–3702 (2017). DOI: https://doi.org/10.1109/CVPR.2017.11

    work page doi:10.1109/cvpr.2017.11 2017

  35. [36]

    Graph Attention Net- works

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Li `o, and Y . Bengio, “Graph Attention Net- works”, inInt. Conf. Learn. Represent.(2018). https://openreview.net/forum?id=rJXMpikCZ

    2018

  36. [38]

    Center -based 3D Object Detection and Tracking,

    J. Sun, Z. Shen, Y . Wang, H. Bao, and X. Zhou, “LoFTR: Detector-Free Local Feature Matching with Transformers”, inProc. IEEE/CVF Conf. Com- put. Vis. Pattern Recognit.8922–8931 (2021). DOI: https://doi.org/10.1109/CVPR46437.2021.00881

    work page doi:10.1109/cvpr46437.2021.00881 2021

  37. [39]

    Effective Approaches to Attention-based Neural Ma- chine Translation

    M.-T. Luong, H. Pham, and C. D. Manning, “Effective Approaches to Attention-based Neural Ma- chine Translation”, inProc. 2015 Conf. Empirical Methods Nat. Lang. Process.1412–1421 (2015). DOI: https://doi.org/10.18653/v1/D15-1166

    work page doi:10.18653/v1/d15-1166 2015

  38. [40]

    Attention Is All You Need

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention Is All You Need”, inAdv. Neural Inf. Process. Syst.30, 5998–6008 (2017). https://papers.nips.cc/paper/7181-attention-is-all-you- need

    2017

  39. [41]

    Continuous gradient fusion class activation mapping: segmentation of laser-induced damage on large-aperture optics in dark-field images

    Y . Han, Y . Huang, H. Dong, F. Chen, F. Zeng, Z. Peng, Q. Zhu, and G. Liu, “Continuous gradient fusion class activation mapping: segmentation of laser-induced damage on large-aperture optics in dark-field images”, High Power Laser Sci. Eng. 12, e4 (2024). DOI: https://doi.org/10.1017/hpl.2023.85

    work page doi:10.1017/hpl.2023.85 2024

  40. [42]

    Random Features for Large- Scale Kernel Machines

    A. Rahimi and B. Recht, “Random Features for Large- Scale Kernel Machines”, inAdv. Neural Inf. Pro- cess. Syst.20 (2007). https://papers.nips.cc/paper/3182- random-features-for-large-scale-kernel-machines

    2007

  41. [43]

    and Mildenhall, Ben and Fridovich-Keil, Sara and Raghavan, Nithin and Singhal, Utkarsh and Ramamoorthi, Ravi and Barron, Jonathan T

    M. Tancik, P. P. Srinivasan, B. Mildenhall, S. Fridovich- Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. T. Barron, and R. Ng, “Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains”, inAdv. Neural Inf. Process. Syst.33 (2020). arXiv: https://arxiv.org/abs/2006.10739

    work page arXiv 2020

  42. [44]

    Sinkhorn Distances: Lightspeed Computation of Optimal Transport

    M. Cuturi, “Sinkhorn Distances: Lightspeed Computation of Optimal Transport”, inAdv. Neural Inf. Process. Syst.26, 2292–2300 (2013). https://papers.nips.cc/paper/4927-sinkhorn-distances- lightspeed-computation-of-optimal-transport

    2013

  43. [45]

    URL https : / / openaccess

    P.-E. Sarlin, D. DeTone, T. Malisiewicz, and A. Rabinovich, “SuperGlue: Learning Feature Matching with Graph Neural Networks”, inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit.4938–4947 (2020). DOI: https://doi.org/10.1109/CVPR42600.2020.00499

    work page doi:10.1109/cvpr42600.2020.00499 2020

  44. [46]

    Best-Buddies Similarity— Robust Template Matching Using Mutual Nearest Neighbors

    S. Oron, T. Dekel, T. Xue, W. T. Freeman, and S. Avidan, “Best-Buddies Similarity— Robust Template Matching Using Mutual Nearest Neighbors”, IEEE Trans. Pattern Anal. Mach. Intell. 40, 1799–1813 (2018). DOI: https://doi.org/10.1109/TPAMI.2017.2737424

    work page doi:10.1109/tpami.2017.2737424 2018

  45. [47]

    On the Variance of the Adaptive Learning Rate and Beyond

    L. Liu, H. Jiang, P. He, W. Chen, X. Liu, J. Gao, and J. Han, “On the Variance of the Adaptive Learning Rate and Beyond”, inInt. Conf. Learn. Represent.(2020). DOI: https://doi.org/10.48550/arXiv.1908.03265

    work page doi:10.48550/arxiv.1908.03265 2020

  46. [48]

    Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartog- raphy

    M. A. Fischler and R. C. Bolles, “Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartog- raphy”, Commun. ACM 24, 381–395 (1981). DOI: https://doi.org/10.1145/358669.358692

    work page doi:10.1145/358669.358692 1981

  47. [49]

    Method for matching online and offline dark- field images of damage sites on large-aperture optical elements

    Y . Han, F. Chen, F. Zeng, C. Lu, Z. Peng, and G. Liu, “Method for matching online and offline dark- field images of damage sites on large-aperture optical elements”, Chinese patent CN113237888B (June 14, 2022)

    2022