Terahertz (THz) imaging has shown great potential in nondestructive testing and biomedical diagnostics due to its non-ionizing and penetrative characteristics, but its practical value remains constrained by low resolution, strong noise, and blurred textures. Existing super-resolution (SR) methods designed for natural images often fail in the THz images, leading to spurious artifacts that obscure defect features. To address these limitations, we propose a Dual-Branch Texture Transfer Super-Resolution (DBTT-SR) framework. The model employs a dual-branch generator, where one branch reconstructs structural components from low-resolution inputs and the other one recovers high-frequency textures from reference images, with cross-attention and deformable convolution enabling dynamic feature alignment. A multi-scale discriminator further imposes hierarchical adversarial supervision, ensuring both local detail fidelity and global structural consistency. Experiments conducted on a dual-frequency THz defect dataset demonstrate that DBTT-SR achieves superior quantitative and qualitative performance, reaching 12.01 dB in PSNR and faithfully preserving defect-specific textures that previous methods fail to recover. These results establish DBTT-SR as an effective and generalizable solution for high-fidelity THz imaging with promising applications in industrial inspection and biomedical imaging.

Imaging, Defect Analysis, Texture Transfer, Super-resolution Reconstruction (SR), Adversarial Learning, Defect Analysis