Loading article...
Loading article...
Stream/Area:
-
Subject:
-
Paper ID:
IJSREM68702
Volume:
10 |
11 |
Nov 2026
Pages:
1-8
Publication Frequency:
Monthly (12 issues per year)
License:International Journal of Scientific Research in Engineering and Management follows the Creative Commons Attribution 4.0 (CC BY 4.0) License, allowing free use and distribution with proper attribution.

Artificial Intelligence And Data Science
Sri Krishna College Of Engineering and Technology
An inspection system is far more useful when the operator can see why an image was flagged. We study Sentinel-X, a ResNet18 classifier for normal versus anomalous products, and ask how well its Grad-CAM explanations agree with real defect annotations on the VisA benchmark. As a classifier it performs well: on 1,629 test images it reaches an AUROC of 0.981 and an accuracy of 95.9%. The heatmaps are a different matter. On 180 anomalous test images (15 per category, all with readable, non-empty masks) the pooled pixel AUROC is 0.756 but the average precision is only 0.030, and at a fixed threshold of 0.5 the mean per-image IoU is 0.020 and the mean Dice coefficient 0.035. A threshold sweep peaks at 0.8 (IoU 0.029, Dice 0.050); because that value was picked on test data, we report it as exploratory. Overlap differs about 28-fold between categories and follows defect size: the thresholded map covers on average 23 times more area than the defect itself, and the categories with the largest defects overlap most. We conclude that Grad-CAM from this classifier shows roughly where the network looked, not where the defect is, and we list the protocol changes needed before such maps could support inspection decisions.
KEYWORDS
Sentinel-X, a ResNet18 classifier, separates normal from anomalous VisA images well (AUROC 0.981, accuracy 95.9% on 1,629 test images), but its Grad-CAM maps localise defects poorly. On 180 anomalous test images with valid masks, pooled pixel AUROC is 0.756 while mean IoU and Dice at τ = 0.5 are only 0.020 and 0.035; overlap differs strongly between categories and follows defect size. Raising the threshold to 0.8 helps on average but hurts several categories, and it was selected on test data. The maps are a coarse guide to the model’s focus and should not be presented as defect segmentations. Future work should (i) choose thresholds on a validation split, per category if possible, and evaluate once on held-out data; (ii) include normal images in a fully specified protocol; (iii) compare against dedicated anomaly-localisation methods such as PatchCore and few-shot DINOv2-based detectors under identical splits and metrics; (iv) report confidence intervals over repeated runs; and (v) analyse overlap by defect size. These steps will show whether the patterns reported here are stable and whether an anomaly-map-based approach is needed for practical inspection.
[1] Y. Zou, J. Jeong, L. Pemula, D. Zhang, and O. Dabeer, “SPot-the-Difference self-supervised pre-training for anomaly detection and segmentation,” in Proc. European Conf. Computer Vision (ECCV), 2022, pp. 392–408, doi: 10.1007/978-3-031-20056-4_23. [2] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778, doi: 10.1109/CVPR.2016.90. [3] R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-CAM: Visual explanations from deep networks via gradient-based localization,” in Proc. IEEE Int. Conf. Computer Vision (ICCV), 2017, pp. 618–626, doi: 10.1109/ICCV.2017.74. [4] P. Bergmann, M. Fauser, D. Sattlegger, and C. Steger, “MVTec AD – A comprehensive real-world dataset for unsupervised anomaly detection,” in Proc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2019, pp. 9584–9592, doi: 10.1109/CVPR.2019.00982. [5] L. Heckler-Kram et al., “The MVTec AD 2 dataset: Advanced scenarios for unsupervised anomaly detection,” Int. J. Computer Vision, 2026, doi: 10.1007/s11263-026-02743-0. [6] T. Defard, A. Setkov, A. Loesch, and R. Audigier, “PaDiM: A patch distribution modeling framework for anomaly detection and localization,” in Pattern Recognition. ICPR International Workshops and Challenges, 2021, doi: 10.1007/978-3-030-68799-1_35. [7] K. Roth, L. Pemula, J. Zepeda, B. Schölkopf, T. Brox, and P. Gehler, “Towards total recall in industrial anomaly detection,” in Proc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2022, doi: 10.1109/CVPR52688.2022.01392. [8] H. Deng and X. Li, “Anomaly detection via reverse distillation from one-class embedding,” in Proc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2022, doi: 10.1109/CVPR52688.2022.00951. [9] K. Batzner, L. Heckler, and R. König, “EfficientAD: Accurate visual anomaly detection at millisecond-level latencies,” in Proc. IEEE/CVF Winter Conf. Applications of Computer Vision (WACV), 2024, doi: 10.1109/WACV57701.2024.00555. [10] J. Jeong, Y. Zou, T. Kim, D. Zhang, A. Ravichandran, and O. Dabeer, “WinCLIP: Zero-/few-shot anomaly classification and segmentation,” in Proc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2023, doi: 10.1109/CVPR52729.2023.01878.
Publication Date
11/5/2026
Peer Review Type
Double-Blind Peer Review
Review Rounds: 2
Documents
Article Views
0
Total views since publication
PDF Downloads
4
Total downloads since publication
Geographic Reach
5
Digital Preservation: This article is permanently archived on Zenodo and complies with all data preservation standards across major indexing platforms.
Plagiarism Screening
This manuscript has been checked for plagiarism using TurnitinPlagiarismChecker software.
Conflict of Interest
The authors declare no conflict of interest related to this research work.
Funding Information
No funding information available for this research.
Related research in the same subject area
Loading similar articles...