An Optimized SwinIR-based Deep Learning Model for Enhanced Image Super-Resolution and Denoising

Image quality enhancement remains a critical challenge in computer vision, particularly in tasks such as super-resolution and denoising, where the balance between fidelity and perceptual realism is essential. Traditional approaches, including GAN-based models like ESRGAN and Real-ESRGAN, have achieved notable improvements but often suffer from texture inconsistencies, artifacts, and limitations in capturing long-range dependencies. To address these gaps, this study proposes an optimized deep learning-based mathematical model for image super-resolution and denoising using SwinIR, a Transformer-driven architecture. By leveraging shifted window-based self-attention, SwinIR effectively models both local and global contextual features, thereby improving reconstruction quality across diverse degradation conditions. The proposed model is extensively evaluated on benchmark datasets such as DIV2K, Set5, and Urban100, considering both full-reference metrics (PSNR, SSIM) and perceptual quality measures (LPIPS, NIQE). Experimental results demonstrate that the SwinIR-based mathematical framework significantly outperforms existing CNN and GAN-based methods in terms of structural accuracy, detail preservation, and noise suppression. Furthermore, the model exhibits robust generalization to real-world low-quality inputs, highlighting its potential for applications in medical imaging, satellite image restoration, and digital photography. This research contributes to advancing deep learning-based mathematical models for image enhancement, offering a scalable and high-performance solution for real-world super-resolution and denoising tasks.