Unified Anomaly Detection via Multi-Scale Contrasted Memory

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep anomaly detection models struggle with edge-case normal samples and varying anomaly scales.
• Current methods lack a unified framework for both unsupervised (UNS) and imbalanced supervised (IMS) anomaly detection settings.

Purpose of the Study:

• To develop a novel, unified framework for deep anomaly detection addressing limitations in UNS and IMS settings.
• To improve robustness against edge-case normal samples and maintain performance across diverse anomaly scales.

Main Methods:

• A two-stage method leveraging multi-scale normal prototypes for anomaly deviation scoring.
• Memory-augmented contrastive learning for joint representation and multi-scale memory module learning.
• An anomaly distance-based detector computing spatial deviation maps using learned prototypes.

Main Results:

• Outperforms state-of-the-art (SoTA) models on various anomalies (object, style, local) and applications (industrial inspection, face anti-spoofing).
• Achieves performance comparable to SoTA out-of-distribution detectors.
• Demonstrates exceptional and consistent performance across both UNS and IMS settings, a first in the field.

Conclusions:

• The proposed method offers a robust and unified solution for deep anomaly detection in both UNS and IMS scenarios.
• The multi-scale prototype approach effectively captures normal data features and adapts to anomaly complexity.
• This work advances anomaly detection capabilities, particularly for challenging edge cases and diverse anomaly types.