AFTD-Net: real-time anchor-free detection network of threat objects for X-ray baggage screening

X-ray baggage screening is a vitally important task to detect all kinds of threat objects at controlled access positions, which can prevent crime and guard personal safety. It is generally performed by screeners to visually determine whether a bag contains threat objects. However, manual detection shows several limitations, from the detection errors to different detection results produced by different screeners. To address these issues, several automated detection approaches have been proposed; nevertheless, none of the methods can achieve end-to-end detection and the results have only classification information without positional information. In this paper, we propose a real-time anchor-free detector of threat objects that can recognize threat objects without using pre-designed anchor boxes. We employ a lightweight but strong backbone network: MobileNetV2 to extract the multi-level information. The backbone network is followed by a deformation layer which aims at handling the nonrigid deformation of threat objects in X-ray images. To further strengthen the proposed network, we design a context enhancement module to aggregate the multi-scale features and generate the enhanced features. We name the network as anchor-free detection network of threat objects (AFTD-Net). We demonstrate the effectiveness of the proposed method against other object detection algorithms on the GDXray database. Our AFTD-Net is a fully convolutional network which does not need any pre-designed anchors and achieves a real-time computation speed of 44.8 FPS.