Detecting prohibited objects with physical size constraint from cluttered X-ray baggage images

X-ray baggage image inspection aims to detect prohibited objects. Existing inspection systems often rely on humans to scrutinize X-ray images. Although several deep-learning-based prohibited object detection methods have been proposed to facilitate human inspection, they often neglect the actual physical sizes of items and thus often lead to many false alarms. To address this issue, in this paper we propose a two-stage prohibited object detection network to identify prohibited objects from heavily cluttered X-ray baggage images. In particular, we first statistically analyzed the physical size distribution of different prohibited object categories and found that their physical sizes exhibit clear distinction. Therefore, we formulated this physical size constraint as a regularization term during the process of training the proposed detection network. Current X-ray datasets only provide annotations of prohibited objects while ignoring uninhibited ones. This may lead to many false positives during testing. Thus, we propose a hard-negative-sample selection scheme to generate proposals of common goods from segmented foreground regions. With these selected hard negative samples, the proposed detector can better distinguish prohibited objects while precluding overfitting on the training dataset. Extensive experimentation demonstrates that the proposed method outperforms state-of-the-art object detection methods. The source code and pre-trained models will be released at https://github.com/xraydetec/Xdet.(C) 2021 Elsevier B.V. All rights reserved.