USTSR: a U-shaped swin transformer framework for super-resolution in solid-state THz array imaging systems

Terahertz (THz) imaging is a promising tool for security screening, biomedical diagnostics, and nondestructive testing, but its effectiveness is severely constrained by limited spatial resolution, noise, and diffraction-induced artifacts. Conventional THz imaging relies predominantly on scanning-based setups, resulting in slow acquisition speeds and low spectral resolution. Here, we present an innovative, integrated imaging approach combining a solid-state frequency multiplier-based focal-plane THz array system-enabling real-time imaging with a dedicated transformer-based super-resolution neural network (USTSR). Without requiring additional complex hardware or extensive computational resources, USTSR significantly enhances image resolution and effectively suppresses noise, tailored specifically for THz image characteristics. Additionally, we introduce an innovative dual-metric evaluation strategy integrating multi-scale structural similarity (MS-SSIM) and Fréchet inception distance (FID), uniquely designed for precise THz A-scan quality assessment. Experiments demonstrate superior performance compared with state-of-the-art methods, achieving 33.17 dB PSNR and 0.86 SSIM. These advancements enable robust and rapid THz imaging with high fidelity, underscoring broad applicability across various practical measurement tasks.