A unified framework for image anomaly detection via reconstruction, segmentation and spatial relationship modeling

Image anomaly detection can safeguard product quality from problems in industrial manufacturing. Existing methods have been successful in detecting structural anomalies, but still exhibit significant limitations in identifying logical anomalies, particularly those involving component quantities and spatial relationships. This paper proposes a unified anomaly detection framework that integrates reconstruction and segmentation-based approaches to comprehensively detect three types of anomalies: structural, component quantity and component relationship anomalies. Specifically, we introduce a spatial relationship memory to capture component layout patterns, a segmentation histogram memory to detect abnormal quantity distributions, and a reconstruction module to localize structural defects. Our method further supports anomaly type interpretation by explicitly indicating the nature of the detected anomaly. Extensive experiments on the MVTec LOCO AD dataset demonstrate that our framework achieves state-of-the-art performance on logical anomaly detection and competitive results on structural anomalies, offering a more interpretable and robust solution.