Aesthetic-Driven Image Enhancement by Adversarial Learning

This is the demo code for the paper Aesthetic-Driven Image Enhancement by Adversarial Learning by Deng et al. at ACMMM 2018. (ArXiv).

The code is based on: [1] https://github.com/soumith/dcgan.torch [2] https://github.com/qassemoquab/stnbhwd

Prerequisite

1. Install torch7 (lua)

   git clone https://github.com/torch/distro.git ~/torch --recursive
   cd ~/torch; bash install-deps;
   ./install.sh;
   source ~/.bashrc;

2. Install the standard STN package

git clone https://github.com/qassemoquab/stnbhwd.git; cd stnbhwd && luarocks make stnbhwd-scm-1.rockspec;

3. Copy the customized STN files to the torch folder path.

   cp ./stn/*.lua ~/torch/install/share/lua/5.1/stn/

4. Other lua dependencies

   luarocks install cudnn; # (The current version requires cudnn 5.0)
   luarocks install tds;
   luarocks intsall nngraph;
   luarocks install dpnn;
   luarocks install matio; # (Make sure that the shared libraries (libmatio.so or libmatio.dylib) are in your library path, e.g., sudo apt-get install libmatio2)
   luarocks install https://raw.githubusercontent.com/szym/display/master/display-scm-0.rockspec

5. Display UI

Optionally, for displaying images during training and generation, we will use the display package. Start the server with: th -ldisplay.start Open this URL in your browser: http://localhost:8000

Download pre-trained models and images:

You may download the models and our images (from AVA dataset) and extract to the repective directories from Google_Drive or Baidu Pan

To run enhancement for an input image:

   cd ./ACMMM_2018_release;
   ./run_single_image.sh

To run training (~11 GB of GPU memory required):

  cd ./ACMMM_2018_release;
  ./train_my_own.sh