A Diffusion-Based Framework with Feature Retrieval for Multi-Class Industrial Anomaly Detection

Mixed-line production with minimal reconfiguration is a critical requirement in flexible manufacturing systems. In this scenario, we propose a Diffusion-based framework with Feature Retrieval (DiFR) for multi-class industrial anomaly detection, it is a unified model to detect anomalies from objects belonging to multiple classes when only normal data is accessible. In DiFR, we first leverage robust image reconstruction capability of diffusion model to mitigate the "identity shortcut" problem. We then introduce a two-stage feature retrieval strategy to replace all foreground features of the input image with their most similar normal counterparts from the feature bank. This helps to prevent the penetration of anomalous information and fundamentally alleviates the "identity shortcut" problem in traditional reconstruction-based approaches. Extensive experiments on benchmark datasets MVTec-AD and MPDD show that the proposed DiFR achieves competitive results compared to the state-of-the-art methods by a significant margin in terms of average AUROC.