SRGAN

• An implement of SRGAN(Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network) for tensorflow version.
• In this repo, vgg19 is not used, instead, MSE is ued to train SRResNet. If you want to use vgg19 to calculate the content loss, you can download model that trained in ImageNet. Then you just need to load to your model during training phase.

Requirements

• tensorflow 1.3.0
• python 2.7.12 or python 3.*
• numpy 1.13.1
• scipy 0.17.0

Usages

downlaod repo

• download this repo by the following instruction:

  $ git clone https://github.com/nnuyi/SRGAN.git
  $ cd SRGAN
  

download datasets

• Firstly, you need to make some directories in the root path(in SRGAN directory)

  $ mkdir data
  $ cd data
  $ mkdir train
  $ mkdir val
  $ mkdir test   
  

train data

• In this repo, I use parts of ImageNet datasets as train data, here you can download the datasets that I used.

• After you have download the datasets, copy ImageNet(here I only use 3137 images) datsets to /data/train, then you have /data/train/ImageNet path, and training images are stored in /data/train/ImageNet

• I crop image into 256*256 resolution, actually you can crop them according to your own.

val data

• Set5 dataset is used as val data, you can download it here.

• After you download Set5, please store it in /data/val/ , then you have /data/val/Set5 path, and val images are stored in /data/val/Set5

test data

• Set14 dataset is used as test data, you can download it here.

• After you download Set14, please store it in /data/test/ , then you have /data/test/Set14 path, and val images are stored in /data/test/Set14

training

  $ python main.py --is_training=True --is_testing=False

testing

  $ python main.py --is_training=False --is_testing=True

Experimental Results

Factor 4(two shuffle layers is used)

low resolution	high resolution GT	high resolution GEN

Factor 4(whole test images)

sampling image
256*256 resolution left:GT right:GEN

References

• Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network
• Is the deconvolution layer the same as a convolutional layer

Contacts

Email:[email protected]