A terahertz image super-resolution reconstruction algorithm based on the deep convolutional neural network

Abstract

Nowadays with the growing threat of terrorist attacks throughout the world, effective security technologies are of urgent need to protect crowds and critical infrastructure. Terahertz wave has emerged as a more powerful tool in security. Terahertz wave is able to penetrate dielectrics such as plastic and cloth so as to detect weapons and contraband hidden under people's clothing without harming human bodies. Nevertheless, image obtained in this frequency range is pretty poor because the diffraction at their relatively long wavelength cannot be ignored in such case. In this paper, we shall briefly introduce the high-resolution (HR) reconstruction for terahertz imaging utilizing the ideology and methodology of super-resolution (SR) restoration in image processing which aims at recovering a high-resolution image from a single low-resolution image. Through the preliminary feasibility research, we applied the image super-resolution algorithm based on the deep convolutional neural network (CNN) to the single passive terahertz image reconstruction. Our deep CNN demonstrates state-of-the-art restoration quality and achieve fast speed as well. Our results indicate that the processed passive terahertz images have clearer edges as well as outlines and are easier to identify suspicious items than the original ones. On the whole, our method outperforms other methods such as the interpolation method and the learning-based image super-resolution reconstruction algorithm. The results indicate a promising prospect for HR terahertz imaging reconstruction.