Wavelet-Guided Dynamic Transformer for High-Fidelity Super-Resolution in Terahertz Imaging Systems

Abstract

Terahertz (THz) imaging is a critical modality for security, biomedicine, and industrial inspection, yet its practical utility is constrained by diffraction limits and low detector array density. Although deep learning has advanced THz image super-resolution (SR), existing models struggle to simultaneously capture multiscale local features and long-range dependencies, leading to the loss of critical high-frequency information. To address these challenges, we propose the wavelet-guided dynamic transformer (WDT), an end-to-end frequency-aware architecture. Specifically, a dynamic attention fusion module (DAFM) adaptively captures multiscale local details and global context through dynamic kernel selection (DKS). Concurrently, a cross-frequency context-aware attention module (CCAM) leverages discrete wavelet transform to decompose features into frequency subbands, supporting efficient cross-frequency fusion while preserving high-frequency components. Extensive experiments on five datasets demonstrate that WDT achieves state-of-the-art performance. Notably, WDT reaches a peak signal-to-noise ratio (PSNR) of 22.71 dB on the THz-CLRM dataset and achieves a 17.3% SSIM improvement over the diffusion-based ResShift on real-world defect datasets. The proposed method significantly enhances edge clarity and structural fidelity, providing a robust computational solution for next-generation THz imaging systems.