Deep Anomaly Detection in Open Worlds

Date

2025

Authors

Ding, Choubo

Editors

Advisors

Shen, Chunhua
Pang, Guansong (Singapore Management University)

Journal Title

Journal ISSN

Volume Title

Type:

Thesis

Citation

Statement of Responsibility

Conference Name

Abstract

Anomaly detection is a critical task in data analysis, with wide-ranging applications across various domains including industrial production, autonomous driving, and medical diagnostics. While traditional methods have long been used to identify deviating data points, they struggle with the increasing volume, dimensionality, and complexity of modern data. This has led to the development of deep learning-based approaches, which offer improved performance through their ability to extract intrinsic features and compute anomaly scores. However, these newer methods also face challenges, particularly with limited labeled anomalies, multiple category distributions, and generalization to unseen anomaly classes. This thesis addresses the challenges of anomaly detection in open-world environments through exploring four interconnected application scenarios. We introduce Disentangled Representations of Abnormalities (DRA) for open-set anomaly detection, effectively utilizing limited labeled anomalous samples. Building on this, we develop Residual Pattern Learning (RPL) for anomaly segmentation, achieving pixel-wise precision in complex, multi-object scenes while maintaining performance on diverse known categories. To address the broader challenge of out-of-distribution detection, we propose Disentangle the Foreground and Background (DFB), which leverages both foreground and background features to enhance detection in complex visual environments. Finally, we present Outlier Label Exposure (OLE) for zero-shot OOD detection, enabling effective identification of anomalies without extensive training data by utilizing auxiliary anomalous category labels and vision-language models. Our comprehensive approach significantly advances the field of anomaly detection, enhancing model adaptability and reducing reliance on large-scale labeled datasets. The proposed methods demonstrate superior performance across various benchmarks and real-world applications. This work not only contributes to the theoretical understanding of anomaly detection but also provides practical solutions for critical applications such as industrial production, autonomous driving, and medical imaging. By addressing the multifaceted challenges of anomaly detection in open-world settings, our research paves the way for more robust, efficient, and interpretable AI systems capable of operating reliably in dynamic, real-world environments.

School/Discipline

School of Computer and Mathematical Sciences

Dissertation Note

Thesis (Ph.D.) -- University of Adelaide, School of Computer and Mathematical Sciences, 2025

Provenance

This electronic version is made publicly available by the University of Adelaide in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. This thesis may incorporate third party material which has been used by the author pursuant to Fair Dealing exceptions. If you are the owner of any included third party copyright material you wish to be removed from this electronic version, please complete the take down form located at: http://www.adelaide.edu.au/legals

Description

Access Status

Rights

License

Grant ID

Published Version

Call number

Persistent link to this record