Yifan Xing



Video Super-Resolution using Generative Adversarial Networks (GAN) with Temporal Fusion

Super-Resolution(SR) aims at recovering a high-resolution image or video from its low-resolution version.It has many applications such as medical imaging, satelliteimaging, and face recognition. In this project, we present a generative adversarial network (GAN) for video super-resolution. To our knowledge, GAN has been applied on single image super-resolution and demonstrated promisingresults, however, there has not been any literacy on videosuper-resolution using GAN. Furthermore, videos have richinformation on the temporal domain which could be bene-ficial to the task of super-resolution. Our target is to builda video super-resolution framework using Generative Ad-versarial Network with temporal information fusion. The effectiveness of this approach is demonstrated through experiments. Comparisons are made against state-of-art per-frame based SRGAN method.

Architecture Overview
Generator Architecture
Discriminator Architecture
Temporal Fusion
Video Demo: Left - Bicubic Interpolation, Middle - Our Method with Slow Temporal Fusion, Right - SRGAN
Check out the repository on GitHub!
Technical Report