Terahertz image enhancement based on a multiscale feature extraction network

Abstract

The development and application of terahertz (THz) waves hold great potential in military, industrial, and biomedical fields. Terahertz time-domain spectroscopy (THz-TDS) imaging systems capture a sample’s time-domain spectral signal to achieve imaging through spectral analysis for intensity and phase information. Challenges in terahertz imaging include spatial diffraction limits, poor image contrast and clarity due to atmospheric water molecule absorption, and Gaussian and impulse noise. This study utilizes a generative adversarial network structure in deep learning models to enhance THz image quality by providing improved denoising and resolution. Through the integration of certain encoder and decoder concepts and introduction of pyramid pooling residual dense block module for feature fusion extraction on low-resolution images, a super-resolution network is designed and employed on selected THz images of deformed metal. Multiple standards are introduced for algorithm performance evaluation. Our experimental results demonstrate that compared with bicubic, super-resolution generative adversarial networks (SRGAN), and residual dense network (RDN) algorithms, our algorithm effectively improves image resolution, and removes noise while preserving high-frequency details without introducing unnecessary high-frequency artifacts.