Direction-aware attention for detail-preserving terahertz image super-resolution

Terahertz (THz) imaging plays a vital role in non-destructive testing, aerospace inspection, and biomedical detection due to its ability to penetrate non-conductive materials. However, the point spread function (PSF) and noise introduce image blur and reduce resolution by suppressing high-frequency details. Convolutional neural networks (CNNs) exhibit insufficient feature representation capabilities under complex degradation conditions. To overcome these challenges, we propose a Residual Detail Enhancement Network (RDENet), a direction-aware architecture integrating directional difference convolutions and hierarchical attention mechanisms. The Residual Detail Enhancement (RDE) module embeds horizontal, vertical, angular, and central gradient priors to restore fine structural details, while the Adaptive Channel–Spatial Attention (ACSA) unit dynamically adjusts channel and spatial responses to highlight critical features. Experimental results demonstrate that RDENet surpasses state-of-the-art methods by 2.37 dB in PSNR and achieves superior performance on public THz and other non-destructive testing tasks, indicating strong potential for practical applications.