AnomNet: A Dual-Stage Centroid Optimization Framework for Unsupervised Anomaly Detection.

Area of Science:

• Industrial AI
• Machine Learning for Manufacturing
• Computer Vision for Quality Control

Background:

• Anomaly detection is crucial for industrial quality and safety.
• Current methods face challenges with multimodal anomalies, domain adaptation, and feature discriminability due to domain gaps.
• Pre-trained models often perform poorly on specific industrial data.

Purpose of the Study:

• To propose AnomNet, a novel deep anomaly detection framework.
• To address limitations in generalization, domain adaptation, and feature discriminability.
• To improve anomaly detection and localization in industrial settings.

Main Methods:

• AnomNet integrates a lightweight feature adapter to bridge domain discrepancies.
• A dual-stage centroid learning strategy is employed for training.
• Stage 1 uses separation and entropy regularization for centroid optimization.
• Stage 2 utilizes centroid-based contrastive learning to refine decision boundaries.

Main Results:

• AnomNet achieved 99.5% image-level AUROC on the MVTec AD dataset.
• The framework reached 98.3% pixel-level AUROC for anomaly localization.
• Demonstrated superior performance and robustness compared to existing methods.

Conclusions:

• AnomNet effectively bridges domain gaps in industrial anomaly detection.
• The proposed framework enhances multi-scale feature discriminability.
• AnomNet offers a robust solution for industrial anomaly detection and localization.