This repo contains implementations of papers

Two methods are being used to perform super resolution

• EDSR + Pixel Shuffling + Perceptual Loss
• Pre-trained GAN (traning Generator and critic seperately and then training both)

• Pytorch>=0.3
• Fastai

$ python super.py

• ImageNet 10% sample
• Celeb A

Requires Fastai==0.7 (super_res_edsr.py)

In this method, We're making use of Pixel shuffle along with EDSR (Removal of BN in skip connection)

Original image

Prediction 

Input image

Prediction 

.

• FastAI>=1.45
• Pytorch >= 1.0

Use crappification from (utils.py)

We use a Unet Generator which has Resnet34/50 (Pretrained) as it's head (modifiable) Generator is trained with low quality images with ground truth being images from original dataset.

When pretraining generator only, here's the output:

Input----------------------------------------------------Prediction------------------------------------------Target

Critic is trained on generator images and original images

Check the path variables carefully

$ python3 super_res_gan.py 

• Perceptual Losses for Real-Time Style Transfer and Super Resolution
• Checkerboard artifact free sub-pixel convolution
• Enhanced Deep Residual Networks for Single Image Super-Resolution
• Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network