Title: Orthogonal LoRA Banks Enable Normality-Preserving Continual Industrial Anomaly Detection

Abstract:

Continual anomaly detection in industrial settings using diffusion models is often hindered by catastrophic forgetting and drift in historical normality priors. While existing continual diffusion techniques rely on replay mechanisms or constrained optimization to retain prior knowledge, they do not explicitly isolate or safeguard category-specific normality priors during sequential adaptation. Although low-rank adaptation (LoRA) offers modular residual updates, conventional LoRA approaches fail to freeze historical normality subspaces and allow new adapters to interfere with previous ones.

To overcome these limitations, we introduce a framework for normality-preserving continual anomaly detection comprising two key components: the History Frozen Orthogonal LoRA Bank (HF-OLB) and the Hierarchical Novelty Adaptive Bank Growth module (HNABG). The HF-OLB freezes the pre-trained U-Net backbone along with the learned LoRA banks, ensuring that new task-specific normality residuals are restricted to the orthogonal complement of historical LoRA subspaces. Meanwhile, HNABG manages layer-dependent residual capacity, expanding the bank only when the residual normality novelty surpasses the expressive limits of the current banks.

We conducted extensive experiments on the MVTec and VisA datasets to validate the proposed method. On the demanding VisA 2x6 setting, our approach attained image and pixel-level A-AUROC scores of 83.6 and 91.8, respectively, with False Match (FM) rates of 3.8 and 3.9. This represents a 3.2-point improvement in pixel-level A-AUROC over the state-of-the-art, alongside a reduction in pixel-level FM by 1.3. These findings confirm that our method successfully preserves historical normality priors across long-horizon continual category sequences.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC