MSG-CapsGAN: Multi-Scale Gradient Capsule GAN for Face Super Resolution

摘要

One of the most useful sub-fields of Super-Resolution (SR) is face SR. Given a Low-Resolution (LR) image of a face, the High-Resolution (HR) counterpart is demanded. However, performing SR task on extremely low resolution images is very challenging due to the image distortion in the HR results. Many deep learning-based SR approaches have intended to solve this issue by using attribute domain information. However, they require more complex data and even additional networks. To simplify this process and yet preserve the precision, a novel Multi-Scale Gradient GAN with Capsule Network as its discriminator is proposed in this paper. MSG-CapsGAN surpassed the state-of-the-art face SR networks in terms of PSNR. This network is a step towards a precise pose invariant SR system.