Intro To GAN

Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network Main target is to reconstruct a super-resolution image or high-resolution image by up-scaling low-resolution images such that texture detail in the reconstructed SR images is not lost.

Steps -

  1. It processes the HR images to get down-sampled LR(Low Resolution) images. Now it has both HR and LR images for the training data set.

  2. Pass the LR images through Generator which up-samples and gives SR(Super Resolution) images.

  3. Then it uses a discriminator to distinguish the HR images and back-propagate the GAN loss to train the discriminator and the generator.

I implemented the paper - Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network