Skip to content
Proba-V Challenge
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.

PROBA-V Super Resolution - Solved with SRGAN

This repository contains my submission code to the challenge Proba-V Super Resolution.

The solution proposed here is a Super Resolution Generative Adversarial Network, as described in this paper.

The architecture is slightly modified from the original paper (mainly for performance concerns).

Changes from Original Paper:

* 16 Residual blocks used, but only 1 upsampling block instead of 2.
* As per suggestion from, we are using UpSampling2D as a
  simple Nearest Neighbour Upsampling instead of SubPixelConvolution.
* The number of discriminator filters were all divided by 2, to shrink the amount of trainable parameters.
* Loss Function: a specefic loss function was defined and used as per the competition scoring:


You will need the following:
Python 3.6
tensorflow 1.12.0
keras 2.2.4
numpy 1.16.1
skimage 0.14.2  (very important to get this specific version)
matplotlib, scipy

For training: Relatively powerfull GPU, this model was trained on an NVIDIA GTX 1080.
It is advised to train this model using a better GPU with more VRAM.


to get the dataset, run this command on your shell:
>> python

* This will also create an "output" directory to store results (images and .h5 files)

How to use:

* clone or download this repository,
* run the "" first,
* then run "" (after adjustments)

Note : Please change the number of epochs and batch size accordingly from within the file. by default, the model
runs for 1000 epochs, with 2 batch size. (It is advised to run this model for more than that!!)

Sample Results:

  Left  : Low Resolution Image (input)
  Middle: Super Resolution Image (prediction)
  Right : High Resolution Image (ground truth)

sample 1 sample 2 sample 3


Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network :

Useful github pages:  (big thanks to luis for providing this repository)

Challenge Website:
You can’t perform that action at this time.