Robust Patch Distribution Modeling for Sensory Anomaly Detection - Archive ouverte HAL Accéder directement au contenu
Pré-Publication, Document De Travail Année : 2023

Robust Patch Distribution Modeling for Sensory Anomaly Detection

Modélisation robuste de la distribution des patchs pour la détection des anomalies sensorielles

Résumé

This paper addresses the challenging issue of unsupervised sensory anomaly detection (AD) in noisy image datasets, which is highly relevant in real-world applications where training images are often noisy and exhibit variability. Current approaches for detecting anomalies in such scenarios are limited and do not account for the noisy nature of the data. To address this limitation, we propose a new method called RPaDiM (Robust Patch Distribution Modeling), which leverages a pretrained CNN and then filter out contaminated patches and computes robust parameters of multivariate Gaussian distributions to summarize the entire training set. Our experiments demonstrate that RPaDiM outperforms existing techniques such as PaDiM and Patchcore in noisy-free settings, and achieves comparable performance with the state-of-the-art SoftPatch method in noisy frameworks. Our results also show that RPaDiM is able to effectively localize anomalies of different sizes, shapes, and positions in the images, making it a promising solution for industrial applications.
Fichier principal
Vignette du fichier
neurips2023.pdf (846.95 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04370178 , version 1 (02-01-2024)

Identifiants

  • HAL Id : hal-04370178 , version 1

Citer

Kamila Kare, Marc Swynghedauw. Robust Patch Distribution Modeling for Sensory Anomaly Detection. 2023. ⟨hal-04370178⟩
26 Consultations
20 Téléchargements

Partager

Gmail Facebook X LinkedIn More