Image Super Resolution using DCGAN

Estimating high-resolution images when given low-resolution images as input.

Single Image Super Resolution (SISR) is a long-standing problem in Computer Vision. Given an input low-resolution (LR) image, the objective is to estimate a corresponding high-resolution (HR) image. Since there is a one-to-many mapping from an LR to possible HR images, this remains a challenging problem in the community.

The current best approaches to the problem are deep learning based, pioneered by the SRCNN paper by Dong et al. At the time of this project, the state-of-the-art benchmarks have been set mostly by GAN-based methods such as the SRGAN by Ledig et al.

This project is my attempt at implementing a simple DCGAN to perform SISR at a scale of 4x.

Some non-cherry-picked outputs from my model. Not bad!

More details and code for this project are available here.