Programmable broad learning system for baggage threat recognition

Detecting illegal and harmful objects in baggage at airports, subways, and bus stations has always been a difficult task that requires intense focus and concentration. Despite recent advances, developing systems for robust autonomous threat recognition still remains a challenge. In this paper, we propose a novel CNN-driven Broad Learning System, dubbed Programmable BLS, for identifying threat objects from the security X-ray scans. The proposed framework first extracts latent features from the input scan utilizing the CNN backbone. Then, the BLS model uses these features to assess whether or not the candidate scan contains the threat items. Unlike existing approaches, the design adaptation of the BLS architecture (within the proposed framework) is fully autonomous, requiring no human efforts to formulate the optimized combination of layers that gives the best classification performance for the given application. This unique design adaption is based on heuristics and greedy searches that measure the relevance of fusing adjacent node pairs in order to improve the overall network performance. Apart from this, across three datasets, namely, GDXray, SIXray, and COMPASS-XP, we rigorously tested the proposed framework on which it outperforms the state-of-the-art by 0.996%, 4.82%, and 0.934%, respectively, in terms of accuracy, and by 3.56%, 1.71%, and 1.30%, respectively, in terms of F1-score.