Pith. sign in

REVIEW 5 minor 47 references

The paper shows that super-resolution reconstructions can look sharper by standard similarity scores while removing the faint evidence a defect detector needs, so inspection-oriented SR must be judged on preserved task evidence and operatin

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 ↗

T0 review · deepseek-v4-flash

2026-08-01 18:03 UTC pith:TYISTTEL

load-bearing objection Careful, predeclared benchmark showing SR reconstruction metrics don't imply detection evidence; the synthetic-only scope limits generalization but the controlled comparison stands.

arxiv 2607.17401 v1 pith:TYISTTEL submitted 2026-07-19 cs.CV

Does Super-Resolution Preserve Defect Evidence? A Low-False-Call Benchmark for Semiconductor Inspection

Shaoliang Yang , Jun Wang This is my paper
classification cs.CV
keywords super-resolutionsemiconductor inspectiondefect detectionlow false-positive rateoperating-point transferreconstruction fidelityweak defectbenchmark
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 paper's central claim is that super-resolution for semiconductor inspection should be evaluated by whether it preserves defect evidence and transfers its operating point, not by how sharp or similar the reconstructed image looks. In a controlled benchmark with one fixed detector and a predeclared low false-positive budget, the two learned super-resolution models achieve the highest structural similarity yet detect fewer defect pixels than bicubic interpolation in every paired repetition. A direct mask-supervised detector recovers substantially more nominal defect evidence, while an illustrative joint model passes independent clean calibration but violates the held-out false-call limit in all ten repetitions. Weak defects remain nearly undetected for every feasible method, and unchanged policies shift sharply on real SEM texture. The conclusion is that reconstruction fidelity alone does not establish inspection utility.

Core claim

The paper's discovery is that reconstruction fidelity and inspection utility are separable and, in this benchmark, opposed. With the same local-residual detector scoring every image transformation at a predeclared low false-positive rate, the trained super-resolution models reach the highest structural similarity yet recover defect-pixel recall roughly one-third of plain bicubic interpolation, losing in every paired repetition. A direct mask-supervised detector shows substantially more nominal defect evidence exists in the low-resolution observation, while an illustrative joint reconstruction/detection model passes independent clean calibration in all ten repetitions and exceeds the held-out

What carries the argument

The load-bearing device is a controlled evaluation design rather than a new architecture. WaferInspectSR-Bench generates paired high- and low-resolution wafer-like images with exact defect, clean-region, and edge masks; one local residual detector is fixed across all reconstruction methods so that any paired recall change is attributable to the image transformation. A three-stage operating-point procedure — temperature fitting on validation, threshold selection on independent no-defect calibration images, and a single held-out evaluation — converts the false-positive limit into a transferable policy. Track A (image transformations under the common detector) and Track B (direct and joint pred

Load-bearing premise

The benchmark's simulated degradation uses hand-picked blur, noise, and downsampling settings rather than measurements from a real scanner; if actual wafer acquisition degrades images differently, the gap between image sharpness and defect recall reported here may not transfer to production lines.

What would settle it

Run the same benchmark on paired real low- and high-resolution wafer inspection images with exact defect masks and no-defect regions, applying the identical fixed detector and predeclared 3e-4 false-positive rule. The paper's central divergence is overturned if a learned super-resolution model both improves structural similarity and recovers more defect pixels than bicubic interpolation at a feasible operating point in most repetitions.

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

If this is right

  • If the benchmark is right, PSNR and SSIM gains from super-resolution cannot be treated as evidence that inspection decisions improve; a sharper reconstruction can reduce defect-pixel recall under a fixed detector.
  • An operating threshold is a policy: calibration success on clean data does not guarantee the same false-call rate on held-out nominal images, as the joint model fails the limit in all ten repetitions.
  • Inspection-oriented super-resolution evaluation should include clean-only controls, weak defects, and an independent clean selection set, with held-out data scored only once after the policy is frozen.
  • A direct mask-supervised detector can recover substantially more nominal defect evidence from the low-resolution image than reconstruction-only methods, showing that the observation itself carries task signal.
  • Weak-defect recall is near zero for every feasible method, so nominal recall improvements do not imply robustness to faint or unseen defect morphologies.

Where Pith is reading between the lines

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

  • The same split-track control logic extends beyond wafers: any imaging pipeline that reconstructs an intermediate image before a rare-event detection step, such as medical scans, aerial surveys, or materials microscopy, risks trading average fidelity for sparse evidence and could reuse this protocol.
  • A natural next experiment, not run in the paper, is task-aware super-resolution in which the reconstruction loss is supplemented with defect and clean-region losses under the same low-false-call rule; the direct-detector result suggests such a model could recover weak defects without breaking false-call control.
  • The external SEM shifts imply that a threshold selected on one background distribution can change behavior drastically on another, which suggests real deployments should revalidate any preselected policy on a sample of production images before use — a risk the paper states but does not quantify.

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

0 major / 5 minor

Summary. The paper introduces WaferInspectSR-Bench, a controlled benchmark for evaluating super-resolution (SR) methods in semiconductor-defect inspection. It separates a Track A experiment, in which image transformations are compared under one fixed local-residual detector, from a Track B experiment, in which direct and joint task-trained predictors are evaluated under a common three-stage operating-policy selection rule. Ten predeclared seed-level repetitions generate synthetic line/space and contact-hole images with exact defect masks, clean controls, weak defects, and a held-out residue morphology; a 4,591-image Carinthia-S external split is used as an unchanged-policy stress test. The headline results are that the two trained reconstruction models achieve higher SSIM than bicubic but lower defect-pixel recall in all ten paired runs, that a direct DeepLabV3 detector recovers substantially more nominal defect evidence at a feasible false-positive rate, that the joint DPU-WaferSR model passes independent clean calibration yet exceeds the held-out FPR limit in all ten runs, and that weak-defect recall is near zero for every feasible method. The paper concludes that reconstruction fidelity and clean-calibration success do not, by themselves, establish inspection utility.

Significance. The benchmark's contribution is primarily methodological, and it is executed with unusual care. Strengths include predeclared seeds and split manifests, a three-stage selection procedure that separates temperature fitting, clean-calibration threshold selection, and held-out evaluation, seed-level paired statistics rather than pixel-pooled inference, exact masks and clean-region definitions, and separate weak/residue/external controls. The code and configuration files are released, which makes the benchmark reproducible. The central claim—that image-similarity metrics alone are insufficient evidence of inspection value—is established by a controlled counterexample in Track A, and the DPU-WaferSR result cleanly demonstrates that clean-calibration feasibility need not transfer to held-out images. The main limitation is the synthetic degradation model of Eq. (4) with hand-set parameters in Table 5 and the pseudo-low-resolution external layer; the authors explicitly acknowledge in Section 9 that paired production optical data would be required for scanner-specific conclusions. I regard this as a scope boundary rather than a flaw, because the paper does not claim fab-scale performance. T

minor comments (5)
  1. [Section 11] The repository URL is given as "github.com/nbbllxx0/W AFERINSPECTSR-BENCH" with an embedded space; it should be "WAFERINSPECTSR-BENCH" (or the actual repository name). Please correct so the availability statement is directly usable.
  2. [Section 7.4 / Table 13] The running text quotes the bicubic external clean-region FPR as 0.00935±0.00057 while Table 13 reports 93.48±5.67 in units of 10^-4. The two representations are consistent, but the text does not state the unit; add a parenthetical so readers are not misled.
  3. [Section 4.3 / Table 5] The headline finding is demonstrated under a single hand-set degradation configuration. A small sensitivity sweep over the blur, noise, and aliasing parameters of Eq. (4) would strengthen the benchmark's generalizability. The paper's conclusion is not contingent on such a sweep, so I do not view this as blocking.
  4. [Section 9] The text says a capacity-matched task-aware SR experiment is "identified as a next step," but Section 9 does not actually describe the next step. Either add a sentence specifying the ablation (same backbone, same training budget, task supervision varied) or correct the cross-reference.
  5. [Figure 9] The per-image recall values (e.g., 0.943 for DeepLabV3) are much higher than the ten-run aggregate 0.198. The caption already states the examples are not included in summaries, but adding the aggregate values on the panel would prevent over-interpretation of a single favorable image.

Circularity Check

0 steps flagged

No significant circularity; the benchmark's claims are measured outcomes under a predeclared holdout policy, not constructed equivalences.

full rationale

The paper's derivation chain does not contain any constructed equivalence between inputs and claimed predictions. In Track A, the detector h is fixed by Eq. (5) while only the image transformation g changes, and the divergence between SSIM and defect recall is a measured paired result across ten seeds (Tables 8–9), not a quantity fitted from the comparison target. In Track B, temperature and candidate thresholds are fit on validation data, the operating point is selected on independent clean calibration images using a predeclared rule (Algorithm 2, Eq. (12)), and held-out partitions are evaluated once without retuning; DPU-WaferSR's held-out infeasibility is an observed transfer failure (Table 11), not an imposed conclusion. The external Carinthia-S experiment is explicitly scoped as a stress test with unchanged policies, not as a paired reconstruction benchmark (Sections 4.4 and 7.4). No load-bearing self-citation, imported uniqueness theorem, or ansatz-smuggled citation appears, and the paper repeatedly disclaims novelty for the general fidelity–utility divergence. Section 9's statement that Eq. (4) is not a calibrated scanner model is a generalization limitation, not a circular step. The central conclusion—reconstruction quality alone does not establish inspection utility—is supported by controlled, predeclared empirical comparisons.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

No new physical entities are introduced. All free parameters are disclosed protocol choices or synthetic degradation settings that shape the benchmark; the authors state these are design settings rather than a fitted scanner model.

free parameters (6)
  • Degradation blur/defocus scales = σ_b=1.2, σ_d=0.7 (k=2)
    Hand-set headline degradation config (Table 5); not fitted to any scanner, so all benchmark numbers are conditional on these choices.
  • Degradation noise magnitudes = shot 0.025, Gaussian 0.020, Poisson peak 80, scan 0.030, aliasing 0.020, contrast U(0.85,1.15)
    Hand-set perturbation scales in Eq. (4); they affect whether weak defects are still present in the low-resolution observation.
  • False-positive target and tolerance = α=3e-4, tolerance=4.5e-4
    Research-analysis rule in Section 3; determines which policies count as feasible.
  • Candidate threshold set = {3,2,1.5,1,0.75,0.5,0.3}×10^-4
    Hand-set candidate targets in Eq. (12) for operating-point selection.
  • Common detector hyperparameters = σ=2.0, Q0.98, final smoothing 0.6
    Fixed scoring rule in Eq. (5); the central Track A comparison depends on this detector choice.
  • DPU loss and prior weights = positive weight 64, clean 2, edge 0.2, risk 0.05, fusion 0.5
    Hand-set weights for the illustrative joint model only; they affect the calibration-transfer demonstration.
axioms (5)
  • domain assumption Synthetic line/space and contact-hole generator with hand-set degradation is representative of real wafer inspection structure.
    Introduced in Section 4; the entire controlled benchmark and headline numbers rest on this premise. The paper admits it is not a calibrated scanner model in Section 9.
  • domain assumption The fixed local-residual detector is a valid measure of preserved defect evidence.
    Eq. (5); used in all Track A comparisons. The paper states it is intentionally transparent, not optimal (Section 5.2).
  • domain assumption Ten seed-level repetitions capture enough pipeline variance for paired conclusions.
    Section 6.5; seed-level dispersion is the inferential unit, but no hierarchical variance decomposition is provided (Section 9).
  • domain assumption Carinthia-S pseudo-low-resolution SEM masks provide meaningful external stress evidence.
    Section 4.4; the external layer is a transfer stress test, not a paired real acquisition benchmark.
  • standard math Standard probability, statistics, and image-processing background.
    Uses standard metrics and paired-comparison reasoning; no novel mathematical claims.

pith-pipeline@v1.3.0-alltime-deepseek · 24637 in / 12585 out tokens · 121740 ms · 2026-08-01T18:03:03.710798+00:00 · methodology

0 comments
read the original abstract

Super-resolution can make inspection images appear sharper without preserving the evidence needed to detect a defect. We study this failure mode with a benchmark that separates reconstruction from detection and evaluates both at a predeclared low false-positive rate. Ten end-to-end repetitions combine independently generated line/space and contact-hole images with model training, calibration, clean controls, weak defects, and a held-out defect morphology. Every reconstruction is scored by the same local residual detector, while direct and jointly trained detectors form a separate comparison track. Reconstruction fidelity and inspection utility diverge: the two learned reconstruction models attain the highest structural similarity yet detect fewer defect pixels than bicubic interpolation in every paired repetition. A direct DeepLabV3 detector reaches $0.1984\pm0.0385$ recall at $0.000174\pm0.000084$ false-positive rate and satisfies the held-out feasibility criterion in all ten repetitions. An illustrative joint model, DPU-WaferSR, passes independent clean calibration but exceeds the held-out limit in all ten repetitions, demonstrating that calibration success does not guarantee transfer. Weak-defect recall remains near zero for every feasible method. Applying the unchanged policies to 4,591 public Carinthia-S masks further reveals large method-dependent shifts on real SEM texture. These results support a simple conclusion: super-resolution for inspection should be judged by preserved task evidence and operating-point transfer, not reconstruction quality alone.

Figures

Figures reproduced from arXiv: 2607.17401 by Jun Wang, Shaoliang Yang.

Figure 1
Figure 1. Figure 1: Why inspection-oriented SR requires task evidence and operating-point transfer. (a) A representative bridge [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Failure modes and representative current outcomes for inspection-oriented SR evaluation. The image panels [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: End-to-end data and protocol path. Exact masks make low-FPR, clean-region, edge, component, and [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Representative line/space and contact-hole samples. Each generated item stores the high-resolution image, [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Illustrative effects of degradation and weak-defect settings. The panels explain the controlled factors; all [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: DPU-WaferSR used as the illustrative joint participant. The compact trunk feeds reconstruction, defect, and [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Separation of fitting, independent clean selection, and held-out evaluation. Training data update weights; [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Reconstruction fidelity versus defect-evidence preservation. Markers show reconstruction-method means, and [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Representative held-out evidence contrasts under preselected seed-7 policies. All panels in a row show the [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Operating-point transfer under the common three-stage rule. Open markers show seed policies and filled [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: External transfer to 4,591 Carinthia-S masks under unchanged synthetic policies. Left: each line pairs one [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Seed-paired operating-point transfer from independent clean calibration to nominal held-out testing. Thin [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] 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

47 extracted references · 3 linked inside Pith

  1. [1]

    Chinmay Belthangady and Loic A. Royer. Applications, promises, and pitfalls of deep learning for fluorescence image reconstruction.Nature Methods, 16(12):1215–1225, 2019

    2019

  2. [2]

    Martin Weigert, Uwe Schmidt, Tobias Boothe, Andreas Müller, Alexandr Dibrov, Akanksha Jain, Benjamin Wilhelm, Deborah Schmidt, Coleman Broaddus, Siân Culley, Mauricio Rocha-Martins, Fabián Segovia-Miranda, Caren Norden, Ricardo Henriques, Marino Zerial, Michele Solimena, Jochen Rink, Pavel Tomancak, Loic Royer, Florian Jug, and Eugene W. Myers. Content-aw...

    2018

  3. [3]

    Evaluation and development of deep neural networks for image super-resolution in optical microscopy.Nature Methods, 18(2):194–202, 2021

    Chang Qiao, Di Li, Yuting Guo, Chong Liu, Tao Jiang, Qionghai Dai, and Dong Li. Evaluation and development of deep neural networks for image super-resolution in optical microscopy.Nature Methods, 18(2):194–202, 2021

    2021

  4. [4]

    Auto- mated semiconductor defect inspection in scanning electron microscope images: A systematic review, 2023

    Thibault Lechien, Enrique Dehaerne, Bappaditya Dey, Victor Blanco, Sandip Halder, Stefan De Gendt, and Wannes Meert. Auto- mated semiconductor defect inspection in scanning electron microscope images: A systematic review, 2023. arXiv:2308.08376

    Pith/arXiv arXiv 2023

  5. [5]

    A multiscale attention mechanism super-resolution confocal microscopy for wafer defect detection.IEEE Transactions on Automation Science and Engineering, 22:1016–1027, 2025

    Xuefeng Sun, Baoyuan Zhang, Yushan Wang, Jialuo Mai, Yuhang Wang, Jiubin Tan, and Weibo Wang. A multiscale attention mechanism super-resolution confocal microscopy for wafer defect detection.IEEE Transactions on Automation Science and Engineering, 22:1016–1027, 2025

    2025

  6. [6]

    AI image enhancement for failure analysis in 3D quantum information technology.Scientific Reports, 15:24078, 2025

    Raphael Wilhelmer, Fabian Laurent, Tatjana Djuric-Rissner, Max Glantschnig, Johann Strasser, Stefan Weinberger, Tobias Herrmann, Clemens Rössler, Peter Czurratis, and Roland Brunner. AI image enhancement for failure analysis in 3D quantum information technology.Scientific Reports, 15:24078, 2025

    2025

  7. [7]

    Leisheng Chen, Kai Meng, Hangying Zhang, Junquan Zhou, and Peihuang Lou. SR-FABNet: Super-resolution branch guided fourier attention detection network for efficient optical inspection of nanoscale wafer defects.Advanced Engineering Informatics, 65:103200, 2025

    2025

  8. [8]

    Super-resolution approach tailored for wafer transmission electron microscopy images.Scientific Reports, 16(1):10662, 2026

    Sungsu Kim, Insung Baek, Hansam Cho, Yongwon Jo, Heejoong Roh, Kyunghye Kim, Munki Jo, Jaeung Tae, and Seoung Bum Kim. Super-resolution approach tailored for wafer transmission electron microscopy images.Scientific Reports, 16(1):10662, 2026

    2026

  9. [9]

    Enhanced deep residual networks for single image super-resolution

    Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. InIEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 1132–1140, 2017

    2017

  10. [10]

    Image super-resolution using very deep residual channel attention networks

    Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. InEuropean Conference on Computer Vision, volume 11211, pages 294–310, 2018

    2018

  11. [11]

    Real-ESRGAN: Training real-world blind super-resolution with pure synthetic data

    Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-ESRGAN: Training real-world blind super-resolution with pure synthetic data. InIEEE/CVF International Conference on Computer Vision Workshops, pages 1905–1914, 2021

    1905

  12. [12]

    SwinIR: Image restoration using swin transformer

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swin transformer. InIEEE/CVF International Conference on Computer Vision Workshops, pages 1833–1844, 2021

    2021

  13. [13]

    Activating more pixels in image super-resolution transformer

    Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image super-resolution transformer. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367–22377, 2023

    2023

  14. [14]

    Simple baselines for image restoration

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. InEuropean Conference on Computer Vision, volume 13667, pages 17–33, 2022. 25 Defect-Evidence Preservation Under Super-ResolutionARXIVPREPRINT

    2022

  15. [15]

    The relationship between precision-recall and roc curves

    Jesse Davis and Mark Goadrich. The relationship between precision-recall and roc curves. InInternational Conference on Machine Learning, pages 233–240, 2006

    2006

  16. [16]

    Carinthia-S dataset, 2025

    Corinna Kofler and Vahidin Hasi´c. Carinthia-S dataset, 2025. Also listed at https://hcai.etf.unsa.ba/datasets.html

    2025

  17. [17]

    Statistical comparisons of classifiers over multiple data sets.Journal of Machine Learning Research, 7:1–30, 2006

    Janez Demšar. Statistical comparisons of classifiers over multiple data sets.Journal of Machine Learning Research, 7:1–30, 2006

    2006

  18. [18]

    Learning a deep convolutional network for image super- resolution

    Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. Learning a deep convolutional network for image super- resolution. InEuropean Conference on Computer Vision, volume 8692, pages 184–199, 2014

    2014

  19. [19]

    Accurate image super-resolution using very deep convolutional networks

    Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. InIEEE Conference on Computer Vision and Pattern Recognition, pages 1646–1654, 2016

    2016

  20. [20]

    Deep laplacian pyramid networks for fast and accurate super-resolution

    Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. InIEEE Conference on Computer Vision and Pattern Recognition, pages 624–632, 2017

    2017

  21. [21]

    Residual dense network for image super-resolution

    Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. InIEEE Conference on Computer Vision and Pattern Recognition, pages 2472–2481, 2018

    2018

  22. [22]

    Pre-trained image processing transformer

    Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12299–12310, 2021

    2021

  23. [23]

    Toward real-world single image super-resolution: A new benchmark and a new model

    Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. InIEEE/CVF International Conference on Computer Vision, pages 3086–3095, 2019

    2019

  24. [24]

    Designing a practical degradation model for deep blind image super-resolution

    Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. InIEEE/CVF International Conference on Computer Vision, pages 4791–4800, 2021

    2021

  25. [25]

    Photo-realistic single image super-resolution using a generative adversarial network

    Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network. InIEEE Conference on Computer Vision and Pattern Recognition, pages 4681–4690, 2017

    2017

  26. [26]

    ESRGAN: Enhanced super-resolution generative adversarial networks

    Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. ESRGAN: Enhanced super-resolution generative adversarial networks. InComputer Vision – ECCV 2018 Workshops, volume 11133, pages 63–79, 2019

    2018

  27. [27]

    Fleet, and Mohammad Norouzi

    Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713–4726, 2023

    2023

  28. [28]

    Bentolila, Comert Kural, and Aydogan Ozcan

    Hongda Wang, Yair Rivenson, Yiyin Jin, Zhensong Wei, Ronald Gao, Harun Günaydın, Laurent A. Bentolila, Comert Kural, and Aydogan Ozcan. Deep learning enables cross-modality super-resolution in fluorescence microscopy.Nature Methods, 16(1):103–110, 2019

    2019

  29. [29]

    Deep learning for sparse scanning electron microscopy

    Patrick Trampert, Sabine Schlabach, Tim Dahmen, and Philipp Slusallek. Deep learning for sparse scanning electron microscopy. Microscopy and Microanalysis, 25(S2):158–159, 2019

    2019

  30. [30]

    Defect Inspection & Review

    KLA Corporation. Defect Inspection & Review. https://www.kla.com/products/chip-manufacturing/ defect-inspection-review. Chip manufacturing product portfolio page; accessed 2026-07-16

    2026

  31. [31]

    KLA BBP 40th Anniversary.https://bbp.kla.com/, 2024

    KLA Corporation. KLA BBP 40th Anniversary.https://bbp.kla.com/, 2024. Accessed 2026-06-01

    2024

  32. [32]

    KLA-Tencor Introduces Comprehensive Wafer Inspection and Review Portfo- lio for Leading IC Technologies

    KLA-Tencor Corporation. KLA-Tencor Introduces Comprehensive Wafer Inspection and Review Portfo- lio for Leading IC Technologies. https://ir.kla.com/news-events/press-releases/detail/121/ kla-tencor-introduces-comprehensive-wafer-inspection-and, 2016. Accessed 2026-06-01

    2016

  33. [33]

    Fully convolutional networks for semantic segmentation

    Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. InIEEE Conference on Computer Vision and Pattern Recognition, pages 3431–3440, 2015

    2015

  34. [34]

    U-Net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention, volume 9351, pages 234–241, 2015

    2015

  35. [35]

    Rethinking atrous convolution for semantic image segmentation, 2017

    Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation, 2017. arXiv:1706.05587

    Pith/arXiv arXiv 2017

  36. [36]

    MVTec AD: A comprehensive real-world dataset for unsupervised anomaly detection

    Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. MVTec AD: A comprehensive real-world dataset for unsupervised anomaly detection. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9592–9600, 2019

    2019

  37. [37]

    The MVTec AD 2 dataset: Advanced scenarios for unsupervised anomaly detection.International Journal of Computer Vision, 134(4):175, 2026

    Lars Heckler-Kram, Jan-Hendrik Neudeck, Ulla Scheler, Rebecca König, and Carsten Steger. The MVTec AD 2 dataset: Advanced scenarios for unsupervised anomaly detection.International Journal of Computer Vision, 134(4):175, 2026

    2026

  38. [38]

    PaDiM: A patch distribution modeling framework for anomaly detection and localization

    Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. PaDiM: A patch distribution modeling framework for anomaly detection and localization. InPattern Recognition. ICPR International Workshops and Challenges, volume 12664, pages 475–489, 2021

    2021

  39. [39]

    Carinthia dataset, 2024

    Corinna Kofler, Sabrina Strauß, Anja Zernig, Ernesto Lazaro Garcia, Michael Boxleitner, Beatrix Mayr, Isabell Dicillia- Kovatsch, and Claudia Anna Dohr. Carinthia dataset, 2024. 26 Defect-Evidence Preservation Under Super-ResolutionARXIVPREPRINT

    2024

  40. [40]

    Task-driven super resolution: Object detection in low-resolution images

    Muhammad Haris, Greg Shakhnarovich, and Norimichi Ukita. Task-driven super resolution: Object detection in low-resolution images. InNeural Information Processing, volume 1516 ofCommunications in Computer and Information Science, pages 387–395. Springer, 2021

    2021

  41. [41]

    Beyond image super-resolution for image recognition with task-driven perceptual loss

    Jaeha Kim, Junghun Oh, and Kyoung Mu Lee. Beyond image super-resolution for image recognition with task-driven perceptual loss. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2651–2661, 2024

    2024

  42. [42]

    Beyond visual fidelity: Benchmarking super-resolution models for large-scale remote sensing imagery via downstream task integration, 2026

    Zhili Li, Kangyang Chai, Zhihao Wang, Xiaowei Jia, Yanhua Li, Gengchen Mai, Sergii Skakun, Dinesh Manocha, and Yiqun Xie. Beyond visual fidelity: Benchmarking super-resolution models for large-scale remote sensing imagery via downstream task integration, 2026. arXiv:2605.00310

    Pith/arXiv arXiv 2026

  43. [43]

    Weinberger

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. InInternational Conference on Machine Learning, pages 1321–1330, 2017

    2017

  44. [44]

    Cooper, and Milos Hauskrecht

    Mahdi Pakdaman Naeini, Gregory F. Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. InAAAI Conference on Artificial Intelligence, volume 29, pages 2901–2907, 2015. AAAI volume 29, issue 1

    2015

  45. [45]

    Goodman.Introduction to Fourier Optics

    Joseph W. Goodman.Introduction to Fourier Optics. Roberts and Company, 3 edition, 2005

    2005

  46. [46]

    Gonzalez and Richard E

    Rafael C. Gonzalez and Richard E. Woods.Digital Image Processing. Pearson Prentice Hall, 3 edition, 2008

    2008

  47. [47]

    Janesick.Scientific Charge-Coupled Devices

    James R. Janesick.Scientific Charge-Coupled Devices. SPIE Press, 2001. 27

    2001