A Comparative Study of Deep Learning Models for Image Super-Resolution Across Various Magnification Levels

Image Super-Resolution (SR) is a vital image processing technique aimed at improving the resolution of lowresolution images to generate high-resolution images, which is essential in applications such as medical imaging, forensic analysis, satellite imagery, etc. Despite its significance, there are still issues with existing SR techniques, such as handling unknown degradation, generalizing across various domains, and recovering the fine details, especially at high scaling factors. Traditional approaches like interpolation and reconstruction-based methods produce poor results, which leads to the use of deep learning models that have demonstrated notable advancements. However, problems like computational complexity, performance drop on real-world data, and dependency on fixed scaling factors remain. In this research, we conduct a comparative analysis of SR models, namely Super-Resolution Convolutional Neural Networks (SRCNN), SwinIR, and Real-ESRGAN, across multiple image domains (e.g., Animals, Cars, Faces, Humans, Landscapes, Ships) using different scaling factors (e.g., <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$x 2, x 3, x 4$</tex>) and metrics like Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Feature Similarity Index (FSIM), Learned Perceptual Image Patch Similarity (LPIPS), Objective Difference Image (ODI) Score, Tenengrad sharpness and Laplacian Variance. This approach helps us better understand the performance of various deep-learning models, guiding further improvements to ensure they are efficient, robust, and applicable to real-world scenarios.