A Cross-domain Vision Transformer Based Framework for Baggage Threat Classification

Abstract

Due to ever-increasing global trade activities and enhanced facilitation in the tourism market, cross-country traveling has seen a massive boost. This has resulted in crowded airports, posing challenges for aviation staff to screen the threat items from passenger baggage. Manual screening of baggage is cumbersome, tiring, and error-prone given the long working hours of the staff combined with concealing strategies incorporated by smugglers to deceive the security system. This has enhanced the requirement of autonomous and robust screening systems at security check-points. Researchers have been working rigorously to develop computer vision-based threat screening systems using different techniques. Recently, transformer-based techniques have been utilized in different classification and localization problems. These algorithms are more effective than traditional machine learning and CNN-based approaches, reducing the errors posed by region-based approaches. In this research, a vision transformer-based cross-domain classification algorithm is introduced for screening baggage threats. The framework uses Vision Transformers architecture and is primarily trained on COMPASS-XP dataset where it outperforms all the previous classification algorithms with an accuracy of 98% and F1-score of 99%. Furthermore, the model showcases the capability of screening threat items from novel datasets by employing small subsets of the corresponding data. Consequently, it exhibits adaptability towards novel image types and is ideal for situations where data scarcity forms the reason for low accuracy.

Keywords Location awareness, Accuracy, Machine learning algorithms, Manuals, Machine learning, Transformers, Real-time systems