SAC-BL: A hypothesis testing framework for unsupervised visual anomaly detection and location.

Area of Science:

• Computer Vision
• Machine Learning
• Statistical Inference

Background:

• Reconstruction-based methods are common for visual anomaly detection (AD), but struggle with weak anomalies.
• Existing AD methods often overlook the crucial role of the decision rule in their statistical framework.

Purpose of the Study:

• To address limitations in current visual anomaly detection methods, particularly concerning weak anomalies and decision rules.
• To introduce a statistically grounded framework for anomaly detection that improves discriminative power.

Main Methods:

• Framed anomaly detection as a multiple hypothesis testing problem.
• Proposed a novel betting-like (BL) procedure integrated with a strong anomaly constraint network (SACNet), termed SAC-BL.
• SACNet is trained to extract discriminative information from weak anomalies, while the BL procedure acts as the decision rule.

Main Results:

• The proposed SAC-BL method demonstrates superior performance in visual anomaly detection compared to existing approaches.
• The framework theoretically guarantees control over the false discovery rate (FDR) at a specified level.
• Extensive experiments validate the effectiveness of SAC-BL, especially in scenarios with challenging, normal-like anomalies.

Conclusions:

• SAC-BL offers a significant advancement in visual anomaly detection by integrating a robust decision rule with a specialized network for weak anomalies.
• The statistical framework provides a principled approach to anomaly detection, enhancing reliability and performance.
• This work opens new avenues for AD research by emphasizing the decision rule and statistical rigor.