Optimal imaging and semantic segmentation of the human body for passive THz security inspection

This paper studies the imaging of a passive terahertz imaging system of a one-dimensional antenna array rotary-scanning system of human body. After singular spectrum analysis, we extract the terahertz energy distribution in the target region for imaging. Then we use the Deeplabv3+ network to semantically segment the human body and carry-ons in the imaging results and use the cycle generative adversarial network to train the terahertz images with the simultaneously acquired optical images to improve the integrity and image resolution of the target contours in the terahertz images. The algorithm is experimentally validated by the passive terahertz imaging system, which can achieve an imaging speed of 4 frames per second and a minimum segmentation accuracy of 3 cm. The algorithm is suitable for passive terahertz imaging of low-sensitivity detectors.