Pith. sign in

REVIEW 4 major objections 6 minor 27 references

CXR-AD: Component X-ray Image Dataset for Industrial Anomaly Detection

T0 review · 4 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read CXR-AD, the first public component X-ray anomaly detection dataset, causes current anomaly detectors to lose about 30 percent average AUROC compared with their MVTec AD results.

desk verdict A useful new dataset benchmark, but the 'publicly accessible' claim is unverifiable without a release mechanism and under-reconciled with prior X-ray data. read the letter →

arxiv 2505.03412 v1 pith:ZLSMRHD4 submitted 2025-05-06 cs.CV

classification cs.CV
keywords X-rayimaginganomalydetectionindustrialdefectsemiconductorcomponentsdatasetbenchmarkpixel-levelannotationlow-contrastmulti-scale
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper introduces CXR-AD, the first publicly available X-ray dataset for anomaly detection in semiconductor components. It is built from real industrial inspection images: 653 normal and 561 defective radiographs across five chip packaging types, with pixel-level masks on every defect. The authors argue that existing datasets cover only surface defects in visible light, so CXR-AD fills a missing regime where defects are internal, low-contrast, noisy, and highly varied in size. To show the regime is genuinely harder, they benchmark a feature-based method, a reconstruction-based method, and a zero-shot multimodal method, reporting an average 29.78% AUROC degradation relative to MVTec AD. If the dataset holds up, it gives the field a realistic testbed for internal defect inspection and a concrete target: algorithms must improve on low-contrast, multi-scale X-ray anomalies.

What carries the argument

The load-bearing object is the dataset itself, built from a YXLON Cheetah microfocus X-ray system and annotated with LabelMe-drawn polygon contours converted to binary masks. The argument is carried by three quantitative characterizations: a grayscale histogram showing that 95.47% of pixels lie between gray levels 50 and 150; a local-contrast statistic $\text{LC} = (\mu_{\text{obj}} - \mu_{\text{bg}})/\mu_{\text{bg}}$ averaging 0.1091 across the UB samples; and a defect-scale analysis in which small defects account for 87% of all defects. These measurements operationalize the claimed challenges of low contrast, high noise, and multi-scale anomalies, and they explain why the benchmarked methods lose performance on CXR-AD; in particular, the paper attributes ONENIP's poor pixel-level showing to boundary blur and leakage of small defects.

What would settle it

Independently re-annotate a random subset of CXR-AD defect images with several trained annotators and compare the new masks against the published ones with IoU and boundary distance; also have an expert verify the geometric-consistency check on the original images. If inter-annotator agreement is low or the automated check misses misaligned contours, the pixel-level benchmark numbers should not be taken at face value.

Watch

Extended reading notes

Core claim

The central claim is that CXR-AD is the first public component-level X-ray anomaly detection benchmark and that it is measurably harder than the de facto surface-defect benchmark MVTec AD. The dataset contains five semiconductor packaging classes (CDIP, CFP16, CFP20, UB, DC); 559 normal images form the training set, and the test set has 94 normal plus 561 defective images with pixel-level polygon masks. Quantitative analysis of the UB class shows 95.47% of pixel values in the 50–150 gray range, an average local contrast of 0.1091 between defect and background, and 87% of defects classed as small; the paper presents these as the mechanisms behind the three named challenges: structure–defect coupling, low contrast with noise, and multi-scale defect morphology. Benchmark results on image/pixel AUROC are PatchCore 56.58/74.78, ONENIP 85.34/93.66, and AdaCLIP 55.32/70.80, versus their MVTec averages of 99.00/98.00, 97.90/97.90, and 89.71/89.90, which the paper summarizes as a 29.78% average degradation.

Load-bearing premise

The pixel-level masks, drawn by a single manual LabelMe workflow and checked only by an automated geometric-consistency module, are assumed accurate enough to serve as ground truth; if they contain systematic errors, the reported pixel-AUROC results and the dataset's quality claims collapse.

Editorial extensions

If this is right

  • CXR-AD provides a public benchmark on which anomaly detection methods can be compared for internal X-ray inspection, alongside surface-defect benchmarks.
  • Because all three method families degrade, progress on CXR-AD requires handling low contrast and small defect scales, not merely scaling up existing feature backbones.
  • The biggest gap is localization: AdaCLIP's pixel AUROC falls to 70.80% and ONENIP's outputs blur defect boundaries, so pixel-level accuracy on X-ray images is the open problem the dataset highlights.
  • The 559-image normal-only training split supports unsupervised and few-shot anomaly detection research without defective training samples.

Reading between the lines

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

  • The 29.78% gap is computed against each method's own MVTec numbers; a fairer difficulty comparison would also control for test-set composition and image resolution, since X-ray images have different statistics than natural-light surface images.
  • A concrete next experiment the paper does not run: apply contrast enhancement or multi-scale feature fusion to the three baselines and measure how much of the MVTec gap closes; that would isolate whether the degradation is inherent to X-ray statistics or to current architectures.
  • The mask-quality assumption could be tested by re-annotating a random subset of CXR-AD defects with multiple annotators and reporting inter-annotator IoU; without that, pixel-AUROC comparisons across future methods may be comparing noise.
  • The 'first public X-ray component anomaly detection dataset' claim is scope-dependent; if adjacent X-ray inspection datasets exist, the distinctive contribution is the combination of semiconductor components, real production defects, and pixel-level masks.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. This paper introduces CXR-AD, an industrial X-ray image dataset for anomaly detection in semiconductor components. It reports 653 normal images (559 train + 94 test normal) and 561 defective test images across five classes, with pixel-level polygonal masks produced with LabelMe. The authors characterize the data using grayscale histograms, local contrast, and defect-scale distributions, and benchmark PatchCore, ONENIP, and AdaCLIP in image- and pixel-level AUROC. They report an average image-level performance degradation of 29.78% relative to MVTec AD, and conclude that existing algorithms are limited for internal defect detection.

Significance. The dataset addresses a genuine gap: most industrial anomaly-detection benchmarks are visible-light surface-defect datasets, whereas X-ray inspection of internal component defects is an important real-world task. The benchmark results are informative, and the acquisition setting (microfocus X-ray system with 16-bit dynamic range) is realistic. The central performance-degradation claim is supported by Table II and is not circular relative to the dataset's descriptive statistics. Provided the dataset is actually released, CXR-AD could become a useful testbed. At present, however, the absence of a public release mechanism makes the 'publicly accessible' claim unverifiable, and several dataset-characterization statistics need correction.

major comments (4)
  1. [Abstract; Section III] The central claim that CXR-AD is the 'first publicly accessible component X-ray anomaly detection dataset' is unsupported because the manuscript contains no repository URL, download link, hosting service, or data-availability statement anywhere in Sections I–V, the footnotes, or the references. Without a release mechanism, the benchmark in Table II cannot be reproduced by other researchers and the claim of public accessibility cannot be verified. The authors should either provide the artifact location or qualify the claim in the abstract and introduction.
  2. [Section II.A] The related-work review does not discuss existing public X-ray defect datasets such as GDXray, which contains X-ray images of castings with defects, and the 'first' claim is not delimited at the point where it is made. If the intended scope is specifically semiconductor packaging components, that restriction should be stated explicitly and the differences from prior X-ray inspection benchmarks should be explained; otherwise the novelty claim is overstated.
  3. [Section III; Figs. 6–8] The dataset-characterization statistics are not internally consistent and are over-generalized. The local-contrast analysis is computed on 62 samples of the UB class only, yet the text concludes that 'the defective regions in semiconductor X-ray images exhibit extremely low contrast characteristics'; the defect-scale statement that small defects account for 87% while medium and large together account for 12% does not sum to 100; and the bin definitions (1%, 10%, 50% of the maximum defect area) are not specified. Please report per-class statistics or clearly state that the analysis is exemplary, correct the percentages, and define the adaptive binning procedure.
  4. [Section III; Table II] The pixel-level mask annotations are created by a single LabelMe-based labeling workflow, with vertex adjustments repeated 5–8 times per sample, and are checked only by an automated geometrical-consistency module. This does not verify the semantic correctness of the defect boundaries, which matters because pixel-level AUROC is one of the paper's main benchmark outputs. The paper should report annotation reliability, for example through a second-annotator study or expert review of a sample of masks, or should explicitly acknowledge this limitation in the conclusion.
minor comments (6)
  1. [Fig. 7] The caption of Fig. 7 repeats 'Distribution of local contrast statistics' from Fig. 6, even though the text describes Fig. 7 as showing the defect-scale distribution; the caption should be corrected.
  2. [Section III] The dataset is called 'the Chip X-ray anomaly detection dataset' at the start of Section III, while the title and abstract use 'Component X-ray'; the terminology should be aligned throughout.
  3. [Abstract; Section IV] The claim of a '29.78% average performance degradation' should specify that this is the mean image-level AUROC degradation across the three methods; the corresponding pixel-level average degradation is different and should be stated separately if used.
  4. [Section IV.C] The sentence 'AdaClip underperforms compared to the other two models, suggesting that it may rely heavily on precise textual prompts for specific detection tasks' is speculative; either provide supporting evidence or rephrase as a hypothesis for future work.
  5. [References] Reference [21] is cited as 'Language models are few-shot learners' but is invoked as 'ChatGPT'; the citation should be corrected to the appropriate GPT/ChatGPT reference, and reference [13] should be checked because the text describes f-AnoGAN rather than the original AnoGAN.
  6. [Section III] The paragraph describing mask conversion contains a duplicated statement about batch-converting JSON annotations to binary masks; the repetition should be removed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmark conclusions are external measurements on a new dataset, not derivations from fitted inputs or self-citations.

full rationale

The paper makes no formal derivation whose output is defined by its inputs. Its central claims are: (i) CXR-AD is a new X-ray anomaly-detection dataset; (ii) the images exhibit low contrast, noise, and multi-scale defects; and (iii) three externally published methods (PatchCore, ONENIP, AdaCLIP) perform worse on CXR-AD than on MVTec AD. Each claim is supported by direct measurement or by running external benchmark methods, not by an equation that forces the conclusion. The dataset statistics (grayscale histogram, local contrast, defect-scale distribution) are descriptive summaries of the images and masks; they are not fitted parameters later renamed as 'challenges' or 'performance degradation.' The 29.78% degradation figure is computed from Table II, which reports independent method evaluations on CXR-AD and MVTec AD, so it is an empirical comparison rather than a construction-derived prediction. There are no self-citations or imported uniqueness theorems; all references are external and none is used to forbid alternatives or to define the dataset's properties. The pixel-level masks are used as ground truth in pixel-AUROC evaluation, but this is standard benchmark practice and does not reduce the performance numbers to the mask-creation procedure. The absence of a download URL or data-availability statement weakens the 'publicly accessible' claim and is a reproducibility/evidence concern, not circularity. No load-bearing step in the paper reduces to its own input by construction, so no circular step can be identified.

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

The paper contains no fitted parameters or invented entities. It relies on domain assumptions that manual pixel annotations are reliable ground truth, that AUROC is an appropriate evaluation metric, and that the single acquisition system used is representative of industrial X-ray inspection.

assumptions (3)
  • domain assumption The manual LabelMe polygon annotations with repeated vertex adjustment constitute accurate pixel-level ground truth.
    Used without inter-annotator agreement or independent expert verification; central to evaluating pixel-level localization.
  • domain assumption Image- and pixel-level AUROC are appropriate metrics for quantifying anomaly detection performance.
    The paper uses these metrics as the sole evaluation, without discussing alternative metrics or cost-sensitive industrial criteria.
  • domain assumption The YXLON Cheetah microfocus X-ray system produces images representative of industrial X-ray inspection for semiconductor components.
    All images come from one machine and one lab; representativeness for other X-ray setups is asserted, not shown.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CXR-AD: Component X-ray Image Dataset for Industrial Anomaly Detection." pith.science (2026). https://pith.science/paper/ZLSMRHD4

@misc{pith2026250503412,
  author       = {Pith},
  title        = {Pith review of: CXR-AD: Component X-ray Image Dataset for Industrial Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZLSMRHD4}},
  note         = {Machine review of arXiv:2505.03412}
}
read the original abstract

Internal defect detection constitutes a critical process in ensuring component quality, for which anomaly detection serves as an effective solution. However, existing anomaly detection datasets predominantly focus on surface defects in visible-light images, lacking publicly available X-ray datasets targeting internal defects in components. To address this gap, we construct the first publicly accessible component X-ray anomaly detection (CXR-AD) dataset, comprising real-world X-ray images. The dataset covers five industrial component categories, including 653 normal samples and 561 defect samples with precise pixel-level mask annotations. We systematically analyze the dataset characteristics and identify three major technical challenges: (1) strong coupling between complex internal structures and defect regions, (2) inherent low contrast and high noise interference in X-ray imaging, and (3) significant variations in defect scales and morphologies. To evaluate dataset complexity, we benchmark three state-of-the-art anomaly detection frameworks (feature-based, reconstruction-based, and zero-shot learning methods). Experimental results demonstrate a 29.78% average performance degradation on CXR-AD compared to MVTec AD, highlighting the limitations of current algorithms in handling internal defect detection tasks. To the best of our knowledge, CXR-AD represents the first publicly available X-ray dataset for component anomaly detection, providing a real-world industrial benchmark to advance algorithm development and enhance precision in internal defect inspection technologies.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 21 canonical work pages

  1. [1]

    YOLO-HMC: An improved method for PCB surface defect detection,

    M. Yuan, Y. Zhou, X. Ren, H. Zhi, J. Zhang, and H. Chen, “YOLO-HMC: An improved method for PCB surface defect detection,” IEEE Transactions on Instrumentation and Measurement, vol. 73, pp. 1–11, 2024

  2. [2]

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

    P. Bergmann, M. Fauser, D. Sattlegger, and C. Steger, “MVTec AD–A comprehensive real-world dataset for unsupervised anomaly detection,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 9592–9600

  3. [3]

    Spot-the- difference self-supervised pre-training for anomaly detection and segmentation,

    Y. Zou, J. Jeong, L. Pemula, D. Zhang, and O. Dabeer, “Spot-the- difference self-supervised pre-training for anomaly detection and segmentation,” in European Conference on Computer Vision, 2022, pp. 392–408

  4. [4]

    Deep learning- based defect detection of metal parts: evaluating current methods in complex conditions,

    S. Jezek, M. Jonak, R. Burget, P. Dvorak, and M. Skotak, “Deep learning- based defect detection of metal parts: evaluating current methods in complex conditions,” in 2021 13th International congress on ultra modern telecommunications and control systems and workshops (ICUMT), 2021, pp. 66–71

  5. [5]

    Generative adversarial nets,

    I. J. Goodfellow et al., “Generative adversarial nets,” Advances in neural information processing systems, vol. 27, 2014

  6. [6]

    Improving unsupervised defect segmentation by applying structural similarity to autoencoders,

    P. Bergmann, S. Löwe, M. Fauser, D. Sattlegger, and C. Steger, “ Improving unsupervised defect segmentation by applying structural similarity to autoencoders,”arXivpreprintarXiv:1807.02011, 2018. 8 > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) <

  7. [7]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “ Image quality assessment: from error visibility to structural similarity, ” IEEE transactionsonimageprocessing, vol. 13, no. 4, pp. 600–612, 2004

  8. [8]

    Fastrecon: Few- shot industrial anomaly detection via fast feature reconstruction,

    Z. Fang, X. Wang, H. Li, J. Liu, Q. Hu, and J. Xiao, “Fastrecon: Few- shot industrial anomaly detection via fast feature reconstruction,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 17481–17490

Show all 27 references
  1. [9]

    Variational Autoencoder for Anomaly Detection: A Comparative Study[J]

    Nguyen H H, Nguyen C N, Dao X T, et al. Variational Autoencoder for Anomaly Detection: A Comparative Study[J]. arXiv preprint arXiv:2408.13561, 2024

  2. [10]

    Learning traces by yourself: Blind image forgery localization via anomaly detection with ViT-VAE,

    T. Chen, B. Li, and J. Zeng, “Learning traces by yourself: Blind image forgery localization via anomaly detection with ViT-VAE,” IEEESignal ProcessingLetters, vol. 30, pp. 150–154, 2023

  3. [11]

    Variational Autoencoder with Gaussian Random Field prior: Application to unsupervised animal detection in aerial images,

    H. Gangloff, M.-T. Pham, L. Courtrai, and S. Lefèvre, “Variational Autoencoder with Gaussian Random Field prior: Application to unsupervised animal detection in aerial images,” ISPRS Journal of PhotogrammetryandRemoteSensing, vol. 218, pp. 600–609, 2024

  4. [12]

    MIAD: A maintenance inspection dataset for unsupervised anomaly detection,

    T. Bao et al., “MIAD: A maintenance inspection dataset for unsupervised anomaly detection,” in Proceedings of the IEEE/CVF internationalconferenceoncomputervision, 2023, pp. 993–1002

  5. [13]

    f-AnoGAN: Fast unsupervised anomaly detection with generative adversarial networks,

    T. Schlegl, P. Seeböck, S. M. Waldstein, G. Langs, and U. Schmidt- Erfurth, “f-AnoGAN: Fast unsupervised anomaly detection with generative adversarial networks,” Medical image analysis, vol. 54, pp. 30–44, 2019

  6. [14]

    Multiscale GAN With Region Adaptive Schemes for Online Inspection of Fabric Flexographic Printing Labels,

    J. Long et al., “Multiscale GAN With Region Adaptive Schemes for Online Inspection of Fabric Flexographic Printing Labels,” IEEE Transactions on Instrumentation and Measurement, vol. 74, pp. 1–18, 2025

  7. [15]

    Few-shot defect image generation via defect-aware feature manipulation,

    Y. Duan, Y. Hong, L. Niu, and L. Zhang, “Few-shot defect image generation via defect-aware feature manipulation,” inProceedingsofthe AAAIconferenceonartificialintelligence, 2023, vol. 37, no. 1, pp. 571– 578

  8. [16]

    Boosting Fine-Grained Visual Anomaly Detection with Coarse-Knowledge-Aware Adversarial Learning,

    Q. Fang, Q. Su, W. Lv, W. Xu, and J. Yu, “Boosting Fine-Grained Visual Anomaly Detection with Coarse-Knowledge-Aware Adversarial Learning,”arXivpreprintarXiv:2412.12850, 2024

  9. [17]

    Semi-Patchcore: A Novel Two-Staged Method for Semi-supervised Anomaly Detection and Localization,

    S. Xie, X. Wu, and M. Y. Wang, “Semi-Patchcore: A Novel Two-Staged Method for Semi-supervised Anomaly Detection and Localization,”IEEETransactionsonInstrumentationandMeasurement, 2025

  10. [18]

    Padim: a patch distribution modeling framework for anomaly detection and localization,

    T. Defard, A. Setkov, A. Loesch, and R. Audigier, “Padim: a patch distribution modeling framework for anomaly detection and localization,” inInternational conferenceonpatternrecognition, 2021, pp. 475–489

  11. [19]

    Towards total recall in industrial anomaly detection,

    K. Roth, L. Pemula, J. Zepeda, B. Schölkopf, T. Brox, and P. Gehler, “Towards total recall in industrial anomaly detection,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 14318–14328

  12. [20]

    ReConPatch: Contrastive patch representation learning for industrial anomaly detection,

    J. Hyun, S. Kim, G. Jeon, S. H. Kim, K. Bae, and B. J. Kang, “ReConPatch: Contrastive patch representation learning for industrial anomaly detection,” inProceedingsoftheIEEE/CVFWinterConference onApplicationsofComputerVision, 2024, pp. 2052–2061

  13. [21]

    Language models are few-shot learners,

    B. Mannetal., “Language models are few-shot learners,”arXivpreprint arXiv:2005.14165, vol. 1, p. 3, 2020

  14. [22]

    Anomalygpt: Detecting industrial anomalies using large vision-language models,

    Z. Gu, B. Zhu, G. Zhu, Y. Chen, M. Tang, and J. Wang, “Anomalygpt: Detecting industrial anomalies using large vision-language models,” in Proceedings of the AAAI conference on artificial intelligence, 2024, vol. 38, no. 3, pp. 1932–1940

  15. [23]

    Matcher: Segment anything with one shot using all-purpose feature matching,

    Y. Liu, M. Zhu, H. Li, H. Chen, X. Wang, and C. Shen, “Matcher: Segment anything with one shot using all-purpose feature matching,” arXiv preprint arXiv:2305.13310, 2023

  16. [24]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab et al., “Dinov2: Learning robust visual features without supervision,”arXivpreprintarXiv:2304.07193, 2023

  17. [25]

    Learning transferable visual models from natural language supervision,

    A. Radford et al., “Learning transferable visual models from natural language supervision,” inInternationalconferenceonmachinelearning, 2021, pp. 8748–8763

  18. [26]

    Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection,

    Y. Cao, J. Zhang, L. Frittoli, Y. Cheng, W. Shen, and G. Boracchi, “Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection,” inEuropeanConferenceonComputerVision, 2024, pp. 55–72

  19. [27]

    Learning to Detect Multi-class Anomalies with Just One Normal Image Prompt,

    B.-B. Gao, “Learning to Detect Multi-class Anomalies with Just One Normal Image Prompt,” in European Conference on Computer Vision, 2024, pp. 454–470

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.