Recognition of anomalies of an a priori unknown type
https://doi.org/10.22405/2226-8383-2022-23-5-227-240
Abstract
In the present article we propose a modification of the PaDiM anomaly detection method which maps images to vectors and then calculates the Mahalanobis distance between such vectors and the distribution of the vectors of the training set. Of the coordinate axes of the vectors we choose a subset of such that the distribution along them is close to normal according to the chosen statistical criterion. The uniformization procedure is then applied to those coordinates and the Mahalanobis distance is calculated. This approach is shown to increase the ROCAUC value in comparison with the PaDiM method.
About the Authors
Alexander Olegovich Ivanov
Lomonosov Moscow State University
Russian Federation
Russian Federation
doctor of physical and mathematical sciences, professor
Gleb Vladimirovich Nosovsky
Lomonosov Moscow State University
Russian Federation
Russian Federation
candidate of physical and mathematical sciences, associate
professor
Vladislav Alexandrovich Kibkalo
Lomonosov Moscow State University; Moscow Center for Fundamental and Applied Mathematics
Russian Federation
Russian Federation
candidate of physical and mathematical sciences
Mikhail Alexandrovich Nikulin
Lomonosov Moscow State University
Russian Federation
Russian Federation
postgraduate student
Fyodor Yurievich Popelensky
Lomonosov Moscow State University
Russian Federation
Russian Federation
candidate of physical and mathematical sciences
Denis Alexandrovich Fedoseev
Lomonosov Moscow State University
Russian Federation
Russian Federation
candidate of physical and mathematical sciences
Ivan Vladimirovich Gribushin
LLC «Huawei Tech Company»
Russian Federation
Russian Federation
lead engineer
Valery Valerievich Zlobin
LLC «Huawei Tech Company»
Russian Federation
Russian Federation
leading engineer of key projects
Sergey Sergeevich Kuzin
LLC «Huawei Tech Company»
Russian Federation
Russian Federation
leading engineer
Ivan Leonidovich Mazurenko
Lomonosov Moscow State University; LLC «Huawei Tech Company»
Russian Federation
Russian Federation
candidate of physical and mathematical sciences, Head of the Laboratory of Intelligent Systems and Data Science HUAWEI
References
1. Falkoner, K. 1990, “Fractal geometry: mathematical foundations and applications”, Chichester
2. etc.: John Wiley & Sons, ISBN 0-471-92287-0, xxii + 288.
3. Defard, T., Setkov, A., Loesch, A., Audigier, R. 2020, “PaDiM: a Patch Distribution Modeling
4. Framework for Anomaly Detection and Localization”, arXiv e-print: https://arxiv.org/pdf/
5. /2011.08785v1.pdf.
6. “Realization of PaDiM approach”, https://github.com/xiahaifeng1995/PaDiM-Anomaly-
7. Detection-Localization-master.
8. Bergmann,P., Batzner,K., Fauser, M., Sattlegger,D., Steger, C. 2021, “The MVTec Anomaly
9. Detection Dataset: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection”,
10. International Journal of Computer Vision, vol. 129, pp. 1038–1059.
11. Venturini, G. M. 2015, “Statistical Distances and Probability Metrics for Multivariate Data,
12. Ensembles and Probability Distributions”, Ph.D. THESIS (advisor: Alberto Munoz Garcia),
13. Dep. of Statistics, Univ. Carlos III, Leganes, Madrid, Spain.
14. Grudic, G. Z., Mulligan, J. 2006, “Outdoor Path Labeling Using Polynomial Mahalanobis
15. Distance”, Robotics: Science and Systems II, August 16-19, 2006, University of Pennsylvania,
16. Philadelphia, Pennsylvania, USA.
17. Shapiro, S. S., Wilk, M. B. 1965, “An analysis of variance test for normality (complete samples)”,
18. Biometrica, vol. 52, pp. 591-611.
19. Ivanov, A., Nosovskiy, G., Chekunov, A., Fedoseev, D., Kibkalo, V., Nikulin, M., Popelenskiy, F.,
20. Komkov, S., Mazurenko, I., Petiushko, A. 2021, “Manifold Hypothesis in Data Analysis: Double
21. Geometrically-Probabilistic Approach to Manifold Dimension Estimation”, arXiv e-print:
22. arXiv:2107.03903 [cs.LG].
23. Stolz, B., Tanner, J., Harrington, H., Nanda, V. 2020, “Geometric anomaly detection in data”,
24. Proceedings of the National Academy of Sciences, vol. 117, pp. 202001741, doi:10.1073/
25. /pnas.2001741117.
26. Erba, V., Gherardi, M., Rotondo,P. 2019, “Intrinsic dimension estimation for locally undersampled
27. data”, Sci. Rep., vol. 9, doi:10.1038/s41598-019-53549-9.
28. Fefferman, C., Mitter, S., Narayanan, H. 2016, “Testing the manifold hypothesis”, J. Amer. Math.
29. Soc., vol. 29, doi:10.1090/jams/852.
30. Bernstein, A., Burnaev, E., Erofeev,P. 2012, “Manifold Reconstruction in Dimension Reduction
31. Problem”, International conference “Intelligent Information Processing” IIP-9.
32. Grassberger,P., Procaccia, I. 1983, “Measuring the strangeness of strange attractors”, Physica
33. D: Nonlinear Phenomena, vol. 9, pp. 189-208.
34. Eckmann, J.-P., Ruelle, D. 1992, “Fundamental limitations for estimating dimensions and
35. Lyapunov exponents in dynamical systems”, Physica D: Nonlinear Phenomena, vol. 56, pp.
36. -187.
37. Mordohai,P., Medioni,G. 2010, “Dimensionality Estimation, Manifold Learning and Function
38. Approximation using Tensor Voting”, Journal of Machine Learning Research, vol. 11, pp. 411-
39.
40. He, J., Jiang, L., Ding, L., Ii, Z. 2014, “Intrinsic Dimensionality Estimation based on Manifold
41. Assumption”, Journal of Visual Communication and Image Representation, vol. 25, issue 5, pp.
42. -747.
43. Granata, D., Carnevale, V. 2016, “Accurate Estimation of the Intrinsic Dimension Using
44. Graph Distances: Unraveling the Geometric Complexity of Datasets”, Sci. Rep., vol. 6,
45. doi:10.1038/srep31377.
46. Levina, E., Bickel,P. J. 2005, “Maximum likelihood estimation of intrinsic dimension”, Advances
47. in neural information processing systems, pp. 777-784.
Review
For citations:
Ivanov A.O., Nosovsky G.V., Kibkalo V.A., Nikulin M.A., Popelensky F.Yu., Fedoseev D.A., Gribushin I.V., Zlobin V.V., Kuzin S.S., Mazurenko I.L. Recognition of anomalies of an a priori unknown type. Chebyshevskii Sbornik. 2022;23(5):227-240. (In Russ.) https://doi.org/10.22405/2226-8383-2022-23-5-227-240
Views:
488
Email this article 