Foreground-aware Prototypical Network for Prohibited Item Detection from X-ray Scans

Abstract

Automatic inspection of X-ray scans is a critical component of modern safety protocols. It plays an indispensable role in detecting concealed weapons, explosives, and other prohibited items that could pose a threat to public safety. Current surveillance systems perform poorly without human intervention. Therefore, many methods have been developed to address this practical problem. However, X-ray scans are complex, with objects often overlapping in a semi-transparent state, which limits the effectiveness of existing methods. To solve the inherent overlapping challenge of X-ray scans, we propose a plug-and-play module called the Foreground-aware Prototypical Network (FaPN). It encourages the model to focus more on prohibited items rather than other irrelevant components. Our proposed FaPN has several excellent properties. First, it is deterministic and discriminative for prohibited items, which is crucial for robust detection in the presence of overlapping. Additionally, it can adapt to multi-scale feature extraction by integrating with existing one-stage detectors such as YOLO series. Comprehensive experiments on datasets verify the effectiveness and efficiency of our model.