Highly Imbalanced Baggage Threat Classification

Recognizing threats in baggage X-ray scans is one of the most crucial tasks for ensuring safety in high-risk areas, including airports, shopping malls, and cargoes, radiograph. Due to the rise in terrorist activity, particularly in the previous two decades, the identification of baggage threats has received the most attention. Nevertheless, this process is time-consuming and restricted by the security officer's inspection capabilities. To overcome this, several frameworks based on deep learning have been suggested to effectively detect contraband items. However, these approaches primarily suffer from the issue of class imbalance, where prohibited objects are rarely seen in the real world compared to harmless baggage content. This paper proposes a novel classification network optimized with the novel compound balanced affinity loss function to address the class imbalance. This proposed loss function is based on the synergic integration of max-margin learning and the effective sample representation. The suggested method is tested on two datasets, COMPASS-XP and SIXray, where it outperforms the state-of-the-art in terms of F1-score by 2.55% and 2.52%, respectively. Also, the proposed approach has surpassed the existing frameworks by attaining accuracy of 89.16% and 70.31%, respectively. To the best of our knowledge, this is the first contour-driven classification framework injected with a compound loss function for highly imbalanced threat classification.