A Comparison of Neural Network-Based Super-Resolution Models on 3D Rendered Images

Super-resolution is an area of Computer Vision comprising various techniques to recover a high-resolution image from a low-resolution counterpart. These techniques can also be used to enhance a low-resolution input image without a native high-resolution original. Single Image Super-Resolution (SISR) techniques aim to do this in a picture-by-picture fashion. In recent years, deep learning models have achieved the best performance, using neural networks to find a mapping between an input low-resolution image and its high-resolution analogous. This work will compare three approaches based on some of the most notable works in neural-network based super-resolution: SRCNN, FSRCNN, and ESRGAN. These methods will be used to enhance 3D computer-generated low-resolution pictures obtained from popular video games and will be evaluated with respect to the quality of the enhanced picture and the required computation time. From our study, we can attest to the superiority of neural network-based methods on the SISR problem and the benefits of taking a perceptual approach when the quality of the resulting images.