superresolution-gan

A DCGAN implementation in Tensorflow for super-resolution of 64 x 64 RGB images to 128 x 128 ones. Related inspiring paper : Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network

Notes on dataset

Worked on the car dataset from ai.stanford.edu. Reason why I preferred this dataset is to maintain the coherence between images. After cropping images according to bounding boxes, eliminated the ones with height < 128 or width < 128, resized them roughly to 128 x 128 x 3 images. Splitting some of them randomly for test, 4D .npys with shape [size, 128, 128, 3] are used in training and test scripts.

Usage

All python scripts have related arguments in order to set hyperparameters, model names, directories etc. and all the information about usage of each script can be seen with --help.

An example usage lies below.

Downloading and Preparing Dataset

Run the script to download dataset:

sh downloadAndExtract.sh

Then run the script,

python3.5 prepare.py

to prepare the dataset before calling,

python3.5 imgs2npys.py

which creates actual training and test sets stored in numpy arrays.

Training and Test

To train a model, run:

python3.5 train.py --model=new_model --save-every=1

This script will train the model new_model and save it under the model/ directory. It also saves model at the end of each epoch (--save-every=1).

To test a trained model, run:

python3.5 test.py --model=new_model

Test script will restore the model from model/ and run on the images residing in test_rgb.npy. Then it will create test output images in test_out_imgs%Y%m%d-%H%M%S/.

Note: superresolution-gan comes with a pre-trained model named as default. Images below are generated with this model.

Example test outputs

Left to right, test output images' format is like this:

• after rough super-resolution with scipy.misc.imresize(img, size=(128, 128, 3)),
• after super-resolution with superresolution-gan,
• original 128 x 128 x 3 image