SUPER RESOLUTION GAN

2 min readNov 27, 2020

We all love to see high resolution images but in reality we don’t see high resolution images all the time. Imagine you have designed a webpage like olx where your users will upload images of their products, we can’t guarantee that all images uploaded will be in good quality, sometimes we have to zoom our images to get more details from it but if we zoom, our image clarity gets distorted. This type disadvantages can be overcomed by Super resolution GAN(SRGAN). Before we get into SRGAN first we will understand what GAN is.

GAN

GAN stands for gnerative adversarial network which was developed by ian j Goodfellow in 2014. It consist of two neural networks which competes with each other and improves their output in each epoches.Those two neural networks are called Generator and Discriminator,Generator use to generate fake data like audio, video,etc to cheat the Discriminator and Discriminator distinguish between real and fake data, in this way both neural networks compete with each other and gets better and better in each epchoes.

TYPES OF GAN

1 Vanilla GAN

2 Conditional GAN

3 Deep Convolutional GAN

4 Laplacian Pyramid GAN

5 Super Resolution GAN

SRGAN

coming back to SRGAN, SRGAN uses GAN system to generate fake images with the help of generator and tries to fool Discriminator and Discriminator tries to differentiate between real data and fake data(generated by Genetrator). Each time when Generator gets caught by Discriminator, it improves its output and make sure that it generates an image which is very close to real data, on the other hand, discriminator improves its performance when it fails to identify the real data. A distorted image is sent as input to Generator and it generates a fake high resolution image which is very close to the real one so that it can fool Discriminator.By this way high resolution image is obtained from low resolution image. After many epoches genrator learns to create a perfect fake HR image from LR image.