Unsupervised Anomaly Instance Segmentation for Baggage Threat Recognition

Identifying potential threats concealed within the baggage is of prime\nconcern for the security staff. Many researchers have developed frameworks that\ncan detect baggage threats from X-ray scans. However, to the best of our\nknowledge, all of these frameworks require extensive training on large-scale\nand well-annotated datasets, which are hard to procure in the real world. This\npaper presents a novel unsupervised anomaly instance segmentation framework\nthat recognizes baggage threats, in X-ray scans, as anomalies without requiring\nany ground truth labels. Furthermore, thanks to its stylization capacity, the\nframework is trained only once, and at the inference stage, it detects and\nextracts contraband items regardless of their scanner specifications. Our\none-staged approach initially learns to reconstruct normal baggage content via\nan encoder-decoder network utilizing a proposed stylization loss function. The\nmodel subsequently identifies the abnormal regions by analyzing the disparities\nwithin the original and the reconstructed scans. The anomalous regions are then\nclustered and post-processed to fit a bounding box for their localization. In\naddition, an optional classifier can also be appended with the proposed\nframework to recognize the categories of these extracted anomalies. A thorough\nevaluation of the proposed system on four public baggage X-ray datasets,\nwithout any re-training, demonstrates that it achieves competitive performance\nas compared to the conventional fully supervised methods (i.e., the mean\naverage precision score of 0.7941 on SIXray, 0.8591 on GDXray, 0.7483 on\nOPIXray, and 0.5439 on COMPASS-XP dataset) while outperforming state-of-the-art\nsemi-supervised and unsupervised baggage threat detection frameworks by 67.37%,\n32.32%, 47.19%, and 45.81% in terms of F1 score across SIXray, GDXray, OPIXray,\nand COMPASS-XP datasets, respectively.\n