Dynamic expandable framework for incremental anomaly detection

Incremental anomaly detection (IAD) has gained significant importance due to the evolving nature of product classes in real-world industrial environments. However, existing IAD methods typically rely on a shared model parameter space, which is prone to catastrophic forgetting when new classes are introduced. To fundamentally address this issue, this paper proposes the Dynamic Expandable Framework (DEF), a novel architecture designed to achieve explicit inter-class decoupling and parameter isolation. The methodology integrates three key components: 1) History-Weighted Feature Selection (HWFS), which selects representative anomaly-sensitive channels to mitigate feature conflicts; 2) a Class-Specific Mixture of Experts (CS-MoE) layer, which assigns dedicated parameters to each class to prevent knowledge overwriting; and 3) the Dynamic Contrastive Routing Network (DCRN), which enables accurate expert selection during inference. Furthermore, t-SNE visualization results qualitatively validate the effectiveness of the proposed decoupling mechanism, demonstrating that DEF significantly mitigates inter-class interference within the decoder layers. Extensive experiments on MVTec-AD and VisA benchmarks show that DEF achieves state-of-the-art (SOTA) performance. Specifically, in the MVTec "3-3 with 4 steps" setting, DEF improves Average Accuracy (ACC) by 4.8% and reduces the Forgetting Measure (FM) by 2.15 compared to current leading approaches. Our approach also maintains superior computational efficiency, requiring only 3.52G FLOPs and 4400MB of training memory, offering an optimal balance for real-world industrial deployment.