Baggage Threat Recognition Using Deep Low-Rank Broad Learning Detector

In the last two decades, monitoring contraband data concealed within baggage has become one of the most pressing security issues. Manual screening of baggage is a time-consuming and error-prone procedure that also compromises the privacy of the passengers. To address this problem, many researchers have proposed X-ray based threat detection models. However, these frameworks involve considerable training efforts on large-scale and well annotated datasets, to accurately detect the prohibited items. This paper presents a novel broad learning detector that is driven via deep low-rank features to identify and localize concealed and cluttered baggage threats from the X-ray imagery. More precisely, the proposed system first extracts the contours of the suspicious baggage items by analyzing the transitional information of the baggage content across multiple orientations. These contours yield a series of proposals which are passed to the CNN backbones to extract distinct features for the objects contained within the proposals. The extracted features are then decomposed through subspace learning and are passed to the broad learning system to identify the respective categories. Moreover, the proposed framework is trained only once using few-shot learning, and it outperforms its competitors by achieving 19.85%, and 8.33% improvements in terms of mean intersection-over-union and mean average precision scores on the highly imbalanced datasets.