Universal Transformation of One-Class Classifiers for Unsupervised Anomaly Detection

TL;DR

A dataset folding method that transforms an arbitrary one-class classifier-based anomaly detector into a fully unsupervised method by making a set of key weak assumptions: that anomalies are uncommon in the training dataset and generally heterogeneous.

Abstract

Detecting anomalies in images and video is an essential task for multiple real-world problems, including industrial inspection, computer-assisted diagnosis, and environmental monitoring. Anomaly detection is typically formulated as a one-class classification problem, where the training data consists solely of nominal values, leaving methods built on this assumption susceptible to training label noise. We present a dataset folding method that transforms an arbitrary one-class classifier-based anomaly detector into a fully unsupervised method. This is achieved by making a set of key weak assumptions: that anomalies are uncommon in the training dataset and generally heterogeneous. These assumptions enable us to utilize multiple independently trained instances of a one-class classifier to filter the training dataset for anomalies. This transformation requires no modifications to the underlying anomaly detector; the only changes are algorithmically selected data subsets used for training. We demonstrate that our method can transform a wide variety of one-class classifier anomaly detectors for both images and videos into unsupervised ones. Our method creates the first unsupervised logical anomaly detectors by transforming existing methods. We also demonstrate that our method achieves state-of-the-art performance for unsupervised anomaly detection on the MVTec AD, ViSA, and MVTec Loco AD datasets. As improvements to one-class classifiers are made, our method directly transfers those improvements to the unsupervised domain, linking the domains.

Figures (21)

• Figure 1: Comparison of prediction on training corruptions of PatchCore with and without folding. Folding was done with 3 votes and 4 folds. Pink outlines are the anomaly ground truth, and regions of increasingly darker blue are predicted anomalies.
• Figure 2: Folding process for removing anomalies from the training dataset for unsupervised anomaly detection with n=2 folds. No assumptions are made about how the OCC anomaly detection model operates or the data type; our method only changes the subset of training data available to the OCC model at each step.
• Figure 3: Example mixture of Gaussians over the predictions by PatchCore on the screw class of MvTecAD for a given fold. Nominal samples are shown in blue, and anomalous samples are shown in orange. The determined crossover threshold between the two fit Gaussians is shown as a red dotted line. The weight of each sample for the mixture of Gaussians was one minus its predicted anomaly value.
• Figure 4: Qualitative results of unsupervised methods on MvTecAD and VisAvisa_data.
• Figure 5: Qualitative results of unsupervised baselines (SoftPatch, FUN-AD), and OCC logical baselines (PUAD, EfficientAD) on Loco AD dataset loco_effecient.