Dynamic expandable framework for incremental anomaly detection.

Area of Science:

• Computer Science, Artificial Intelligence, Machine Learning

Background:

• Incremental anomaly detection (IAD) is crucial for evolving industrial product classes.
• Existing IAD methods suffer from catastrophic forgetting due to shared model parameters.
• Need for methods that explicitly decouple and isolate parameters for new classes.

Purpose of the Study:

• To propose the Dynamic Expandable Framework (DEF), a novel architecture for IAD.
• To achieve explicit inter-class decoupling and parameter isolation.
• To mitigate catastrophic forgetting when introducing new product classes.

Main Methods:

• History-Weighted Feature Selection (HWFS) to select anomaly-sensitive channels and reduce feature conflicts.
• Class-Specific Mixture of Experts (CS-MoE) layer for dedicated parameters per class, preventing knowledge overwriting.
• Dynamic Contrastive Routing Network (DCRN) for accurate expert selection during inference.

Main Results:

• DEF effectively mitigates inter-class interference, validated by t-SNE visualizations.
• Achieved state-of-the-art (SOTA) performance on MVTec-AD and VisA benchmarks.
• Improved Average Accuracy (ACC) by 4.8% and reduced Forgetting Measure (FM) by 2.15 in a specific MVTec setting.
• Demonstrated superior computational efficiency (3.52G FLOPs, 4400MB training memory).

Conclusions:

• DEF successfully addresses catastrophic forgetting in IAD through inter-class decoupling and parameter isolation.
• The proposed framework offers a robust and efficient solution for real-world industrial anomaly detection.
• DEF provides an optimal balance of performance and computational cost for practical deployment.