Unsupervised anomaly instance segmentation for baggage threat recognition

Identifying potential threats concealed within the baggage is of prime concern for security staff. Many researchers have developed frameworks that can automatically detect from X-ray scans. However, to best our knowledge, all these require extensive training efforts on large-scale and well-annotated datasets, which are hard procure in real world, especially rarely seen contraband items. This paper presents a novel unsupervised anomaly instance segmentation framework recognizes threats, scans, as anomalies without requiring any ground truth labels. Furthermore, thanks its stylization capacity, trained only once, at inference stage, it detects extracts items regardless their scanner specifications. Our one-staged approach initially learns reconstruct normal content via an encoder–decoder network utilizing proposed loss function. The model subsequently identifies abnormal regions by analyzing disparities original reconstructed anomalous then clustered post-processed fit bounding box localization. In addition, optional classifier also be appended with recognize categories extracted anomalies. A thorough evaluation system four public re-training, demonstrates achieves competitive performance compared conventional fully supervised methods (i.e., mean average precision score 0.7941 SIXray, 0.8591 GDXray, 0.7483 OPIXray, 0.5439 COMPASS-XP dataset) while outperforming state-of-the-art semi-supervised threat detection 67.37%, 32.32%, 47.19%, 45.81% terms F1 across respectively.