A Regularization Factor-Based Approach to Anomaly Detection Using Contrastive Learning

Anomaly detection methods are introduced to remove abnormalities from specific datasets. In the paper, mean shifted contrastive loss for anomaly detection, a new objective function named mean shifted contrastive loss (MSCL), was created to retain the pre-trained features over transfer learning which...

Full description

Saved in:
Bibliographic Details
Published in:Arabian journal for science and engineering (2011) 2024-03, Vol.49 (3), p.3371-3381
Main Authors: Maurya, Abhinav, Yadav, Mainejar, Yadav, Geetanjali, Saxena, Peehu, Giri, Saksham
Format: Article
Language:English
Subjects:
Abnormalities
Anomalies
Datasets
Engineering
Humanities and Social Sciences
Learning
multidisciplinary
Regularization
Research Article-Computer Engineering and Computer Science
Science
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Anomaly detection methods are introduced to remove abnormalities from specific datasets. In the paper, mean shifted contrastive loss for anomaly detection, a new objective function named mean shifted contrastive loss (MSCL), was created to retain the pre-trained features over transfer learning which were lost in previous anomaly detection methods. MSCL stops learning after a certain epoch as the points get closer to the normalized center. The proposed work investigated that due to some noise in the extracted features from images, the points revolve around the normalized center, and the accuracy freezes after some epochs. To resolve this issue, a new loss function is proposed which has better performance. The proposed methodology obtains a highly effective AUROC score of 98.6% and 97.6% on the CIFAR-100 and CIFAR-10 datasets.
ISSN:2193-567X
1319-8025
2191-4281
DOI:10.1007/s13369-023-07959-7