Gaussian Image Anomaly Detection with Greedy Eigencomponent Selection - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2023

Gaussian Image Anomaly Detection with Greedy Eigencomponent Selection

Tetiana Gula
  • Fonction : Auteur
  • PersonId : 1319190
Joao P C Bertoldo

Résumé

This paper addresses the challenge of Anomaly detection (AD) in images by proposing a novel dimensionality reduction technique using pre-trained convolutional neural network (CNN) with EfficientNet model. We introduce two tree search methods with a greedy strategy for improved eigencomponent selection. We conducted three experiments to evaluate our approach: examining components choice on test set performance when intentionally overfitting, training on one anomaly type and testing on others, and examining training with a minimal image set based on anomaly types. Unlike traditional methods that emphasize variance, our focus is on maximizing performance and understanding component behavior in diverse settings. Results show our technique outperforms both Principal Component Analysis (PCA) and Negated PCA (NPCA), suggesting a promising advancement in AD efficiency and effectiveness.

Dates et versions

hal-04319501 , version 1 (02-12-2023)

Licence

Paternité - Partage selon les Conditions Initiales

Identifiants

Citer

Tetiana Gula, Joao P C Bertoldo. Gaussian Image Anomaly Detection with Greedy Eigencomponent Selection. IEEE/CVF International Conference on Computer Vision, Oct 2023, Paris, France. ⟨hal-04319501⟩
21 Consultations
0 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More