Balanced Transformer for Highly Imbalanced Baggage Threat Recognition

Manual baggage screening, a routine security practice in ensuring the safety of travelers in airports, has several pitfalls. Consequently, researchers have embraced deep learning algorithms to deliver better solutions for the autonomous detection of baggage threats. Nevertheless, these approaches primarily suffer from the class imbalance problem and fail to capture the global context of the threat items. Hence, this paper proposes an abnormality contour-driven classification approach based on visual transformers to model meaningful and distinctive long-range representations from object contours within baggage imagery. Moreover, injecting the framework with the proposed balanced focal loss enables to learn discriminative features of normal and threat objects based on the effective number of samples. We tested the proposed system on two highly skewed public baggage X-ray datasets, where it surpassed state-of-the-art methods by attaining 97.4%, 87.2%, and 97.2%, 88.9% in terms of accuracy, and F1-score, respectively.