Concealed Threat Detection based on Multi-view Millimeter Wave Imaging for Human Body

Near-range millimeter wave (MMW) has attracted much more attentions for its great potential in security check. In this paper, deep-learning based concealed threat detection is researched by MMW human body imaging systems. To improve the detection performance for small objects in MMW images in noises and textures, we design a voting classifier based on multi-view images in a scan. Faster region with conventional neural network (Faster-RCNN) and online hard example mining (OHEM) architectures are utilized to detect objects in each image, where the results are combined and voted for a more reliable detection in the scan. The proposed method is validated on our dataset of 700k MMW human body images, which surpassed the state-of-the-art object detectors both in precision and recall.