Addressing Class Imbalance in X-ray Threat Detection with Self-Supervised Balanced DINO

Aviation security, which is mainly tasked with the assurance of safety to the travelers by offering reliable baggage screening, remains a very critical issue. Traditional manual inspections are prone to errors and labor intensive. This calls for automatic detection systems. However, methods commonly used for X-ray threat detection have the problem of data imbalance and poor feature representation, especially for rare threat items. It proposes a novel Balanced DINO framework that incorporates the Focal-Augmented Loss to tackle the above challenges into a DINO-based Vision Transformer. Our study aims at balancing class representations, using both algorithmic weighting and data augmentation methods, to improve the detection accuracy of the rare threat items. Balanced DINO demonstrates extensive testing on three public datasets, namely SIXray, CLCXray, and COMPASS-XP, with considerably improved performance compared to state-of-the-art methods. Concrete, our framework outperforms the competitors by at most 3.77% in F1-score and at most 2.31% in accuracy on the highly imbalanced SIXray dataset, which firmly proves that our framework is robust and adaptable for different settings.