Multi-scale prototype fusion network for industrial product surface anomaly detection and localization

In complex industrial application scenarios, abnormal samples are scarce. In the case of weak defect features, the high similarity of positive and negative samples further complicates detection and localization. In addition, anomalies are often subtle and unpredictable, which makes it particularly difficult to detect and localize anomalous subregions with unknown anomaly patterns. Many detection algorithms suffer from high computational complexity and huge memory consumption. To address these challenges, this paper proposes a multi-scale prototype fusion network for industrial product surface anomaly detection and localization (MPFnet). MPFnet uses multi-scale prototypes to construct representative normal patterns and incorporates a multi-scale fusion block to facilitate information exchange between different scales. This design enhances the model’s attention to characterize prototype and normal features. Feature adapter is constructed to generate fitness features, reducing domain bias. By adding noise to the adapted features, anomalous features are generated, and anomalies are detected using a simple and efficient discriminator. A large number of experiments were carried out on the challenging MVTec AD and MVTec LOCO AD datasets, demonstrating that MPFnet outperforms other state-of-the-art comparative methods, achieving good detection and localization results regardless of defect patterns.