Programmable Broad Learning System to Detect Concealed and Imbalanced Baggage Threats

Abstract

Manual screening of baggage at airports, shopping malls, and shipments to identify potentially dangerous items is a time-consuming process that requires the unwavering efforts of a human observer. Numerous researchers have addressed this issue by developing autonomous threat detection systems. However, the performance of these systems is still vulnerable to high occlusion and unbalanced contraband data. In this paper, we present a novel programmable CNN-driven broad learning system (BLS) that automatically adapts its design specifications to effectively recognize the concealed and imbalanced contraband data depicted within the baggage X-ray scans. First, the input scan is passed to the CNN backbone to extract distinct latent features. These features are then passed to the BLS model, which determines whether the scan contains potentially dangerous items or not. Additionally, the BLS’s architecture (within the proposed framework) is programmed in such a way that no human effort is required to optimize it for producing the best threat detection performance. This novel design adaptation is performed via heuristics and greedy searches that quantify the importance of each edge fusing the adjacent node pairs to optimize the network’s overall performance. The proposed system is thoroughly tested on three datasets, namely GDXray, SIXray, and COMPASS-XP, on which it leads the state-of-the-art by 2.94%, 19.33%, and 13.38%, respectively, in terms of F1 score.