A comparative study of deep learning models for image super-resolution

Many deep learning-based image super-resolution models exist to effectively up-sample images, with the most notable and reliable architectures being Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and Generative Adversarial Networks (GANs). To date, model benchmarking has been made only with the same architecture type or only with certain datasets that could potentially be beneficial to the proposed models. In this paper, we present the first-known comparison of state-of-the-art super-resolution models, namely, SwinIR, EDSR, Swin2SR and Real- ESRGAN, to serve as a reference baseline for future applications where the modelling complexity, frame rates and overall super-resolution accuracy is of concern. The experiments were conducted by reproducing the models entirely by following the training procedures highlighted in their original paper. Then, we performed the evaluations on the conventional image super-resolution test sets, namely, Set5, Set14, BSD100, Urban100, T91 and Manga109. Our experimental results show that each model has their respective tradeoff between the number of measures taken to suppress the super-resolution artifacts and achieve a higher super-resolution accuracy and the overall model processing times, such as the model convergence speed and their respective frame rates.