<p>Terahertz imaging technology has significant potential across diverse applications, such as security inspection, biomedical research, and non-destructive material testing. Recent advances in deep learning have enabled high-quality restoration of terahertz images, but existing methods often require long training times and complex architectures. Moreover, many of these methods fail to effectively leverage multi-scale features, either neglecting them or applying them at a static scale, which limits their performance in super-resolution tasks. To address these challenges, propose a novel Multi-Dimensional Attention Fusion Network (MAFN) integrated with the Wombat Swin Transform Network (WSTN) for super-resolution enhancement of terahertz images. The MAFN efficiently reduces noise while preserving key details and structural features, which is a critical step in terahertz image restoration. Following denoising, the WSTN further enhances image resolution and quality by mapping low-resolution (LR) images to high-resolution (HR) images through three key modules, shallow feature extraction, deep feature mapping, and reconstruction. To optimize performance, the Wombat optimization algorithm is employed to fine-tune the learning rate for the Swin Transform, maximizing the Peak Signal-to-Noise Ratio (PSNR). The performance of the proposed model was thoroughly evaluated using several metrics, yielding impressive results PSNR of 46.9, MSE of 0.009, CC of 0.99, SSIM of 0.91, QNR of 0.9, and SNR of 22.5. These outcomes demonstrate the superiority of this method over existing techniques, highlighting its ability to achieve high-quality super-resolution in terahertz images. Overall, the integration of MAFN with WSTN and the Wombat optimizer significantly improves the effectiveness of terahertz image restoration and super-resolution.</p>