Terahertz image super-resolution restoration using a hybrid-Transformer-based generative adversarial network

Abstract

Terahertz (THz) imaging and detection technology has been widely used in subway stations, high-speed rail stations, airports, and other security detectors because of its ability to penetrate non-metallic materials such as clothing and paper to detect hidden objects without radiation hazards. However, due to the influence of THz wavelength, optical equipment, particle scattering, and water vapor absorption in the air, the THz images obtained by the existing THz imaging systems often have low imaging resolution and noise interference problems. To solve these problems, we propose a hybrid Transformer-based generative adversarial network (HTSRGAN) to achieve THz image super-resolution (SR) restoration. A generator network is designed to balance the noise removal and the critical context feature information extraction of THz images. A hybrid residual transformer enhancement block (HRTEB) is designed to filter noise and enhance extract information. HRTEB is composed of Residual Spatial and Channel Reconstruction Convolution (SCConv) Enhance Dense (RSED) blocks and the residual Swin Transformer (RSformer) module. To improve the context relevance and robustness of the feature information in the image reconstruction module, we develop an improved LeWinformer (ILformer) module that can stabilize and enhance the information of the target item after upsampling. The experimental results show that the proposed method achieves high-quality THz image SR restoration and performs well on noise elimination, demonstrating better than state-of-the-art comparison methods. The proposed method has potential applications in public security inspection, medical diagnostic imaging, cultural heritage protection, and so on.