MECAD: A multi-expert architecture for continual anomaly detection

TL;DR

MECAD tackles the challenge of continual anomaly detection in industrial environments where object classes evolve over time. It introduces a multi-expert architecture with similarity-driven expert assignment, patch-embedding memory-banks, and replay-based training to enable incremental learning without full retraining. The approach demonstrates that a 5-expert configuration achieves a near-peak AUROC of 0.8259 on MVTec AD while substantially reducing forgetting and optimizing memory usage. This work offers a practical, scalable solution for adapting anomaly detectors to changing product types in real-world factories.

Abstract

In this paper we propose MECAD, a novel approach for continual anomaly detection using a multi-expert architecture. Our system dynamically assigns experts to object classes based on feature similarity and employs efficient memory management to preserve the knowledge of previously seen classes. By leveraging an optimized coreset selection and a specialized replay buffer mechanism, we enable incremental learning without requiring full model retraining. Our experimental evaluation on the MVTec AD dataset demonstrates that the optimal 5-expert configuration achieves an average AUROC of 0.8259 across 15 diverse object categories while significantly reducing knowledge degradation compared to single-expert approaches. This framework balances computational efficiency, specialized knowledge retention, and adaptability, making it well-suited for industrial environments with evolving product types.

Figures (5)

• Figure 1: Overview of the MECAD pipeline. The system processes object classes sequentially, extracting patch-level embeddings from normal samples, Coreset of these embeddings are assigned to specialized experts based on cosine similarity. Each expert maintains a memory buffer with randomly selected samples and is updated only when assigned a new class. During inference, test samples are processed through the corresponding expert to generate anomaly scores at the image level.
• Figure 2: Per-Class Auroc Across Expert Configurations
• Figure 3: Impact of varying the number of experts (1-8) on (a) forgetting mitigation and (b) anomaly detection performance across all classes
• Figure 4: Final evaluation results showing (a) detailed per-class AUROC performance and (b) average AUROC scores achieved by each expert in the 5-expert configuration
• Figure 5: Evolution of (a) forgetting and (b) AUROC metrics for each expert in the 5-expert configuration as classes are introduced sequentially