YOLO-MSFG: Toward Real-Time Detection of Concealed Objects in Passive Terahertz Images

Passive terahertz (THz) systems can perform the perspective imaging of objects concealed underneath clothing on the human body without a radiometric emitter. In this study, a linear-array passive THz imaging system was designed to realize highly efficient push-broom imaging. However, the acquired THz images are contaminated by severe striping and random noise, which greatly hampers object detection. Therefore, this paper aims to report a dedicated deep learning network for the automatic, accurate, and real-time detection of concealed objects in passive THz images. The main work of this research is threefold. First, a bilateral filter is integrated into a convolutional neural network (CNN) to generate a space-range tunable learning architecture. Second, a space-range grid combined with a geometric transformation matrix is designed to provide a core unit for multi-scale filtering and geometric (MSFG) augmentation. Finally, the one-stage YOLOv5 detector is improved to YOLOv5I by refining the granularity of its prediction units. The proposed approach was tested on onsite passive THz images acquired from various scenes, and the results demonstrate that the proposed MSFG augmentation method can significantly improve the detection accuracy of all the candidate CNNs on the passive THz images. Among them, the YOLOv5I-MSFG was found to exhibit the best performance via a comprehensive assessment of its real-time ability and detection accuracy; the time consumption per frame was found to be
$\sim 53$
ms on the configured platform, and the accuracies of one-class and two-class detection were found to reach 93.49% and 90.76%, respectively.