TOWARDS ROBUST HUMAN MILLIMETER WAVE IMAGING INSPECTION SYSTEM IN REAL TIME WITH DEEP LEARNING

With the ever-growing requirements of human security check in public, near-field millimeter wave (MMW) imaging techniques have been developing rapidly in recent years. Due to the lack of MMW images, low resolution and indistinguishable texture in most MMW images, it is still a great challenge to do high performance object detection task on MMW images. In this paper, we propose a novel framework to automatically detect concealed weapons and potential dangerous objects based on a single human millimeter wave image, in which a deep convolutional neural network (CNN) is presented to simultaneously extract features, detect suspicious objects, and give the confidence score. Unlike traditional optical image level solutions, we comprehensively analyze the original MMW data for object representation, incorporate domain-specific knowledge to design and train our network. Moreover, combined with the modern focal loss theory, we devise an effective loss function elaborately to optimize our model. Experimental results on both our dataset and real world data show the effectiveness and improvement of our method compared with the state-of-the-arts.