Unsupervised Threshold Learning with "$L$"-trend Prior For Visual Anomaly Detection

Zhonghang LIU, Jian Ming Ng, Wen Yan LIN

Keywords: Unsupervised, visual anomaly detection, threshold learning

Abstract: This paper considers unsupervised threshold learning, a practical yet under-researched module of anomaly detection (AD) for image data. AD comprises two separate modules: score generation and threshold learning. Most existing studies are more curious about the first part. It is often assumed that if the scoring module is good, estimating an accurate threshold is within easy reach. However, we argue that in the context of computer vision, some challenges in high-dimensional space lead threshold estimation be a non-trivial problem. In this paper, we leverage the inherent difference between normal instances and anomalies by ranking their anomaly score, which shows a phenomenon that involves two distinct trends. We term it as the "$L$"-trend prior. With that finding, we utilize an adaptive polynomial regression model to determine the threshold. Unlike the classic threshold learners which rely on enough training samples or statistical assumptions, this method is plug-and-play that can be implemented into different anomaly score function among various datasets. Also, the evaluation results demonstrate an obvious improvement.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Unsupervised and Self-supervised learning

TL;DR: Propose a new perspective for unsupervised visual anomaly detection