In this work we present a model to automatically colorize black and white images with zero human interference. Our objective is not to match the ground truth, but rather to generate plausible color images that can aesthetically fool a real human observer. Our attempts include implementing a Conditional Deep Convolutional Generative Adversarial Network as well as using the technique of transfer learning to infuse global features into the network from a pre-trained model . Our model is trained on publicly available dataset such as CIFAR-10 and Places365.
Environment:
- Python 2.7
- Ubuntu 16.04
- cuDNN v7.1.4 Runtime Library for Ubuntu16.04 (Deb)
- cuDNN v7.1.4 Developer Library for Ubuntu16.04 (Deb)
Packages:
sudo apt-get install python-pip python-dev
sudo pip install tensorflow
sudo pip install numpy
sudo apt-get install libatlas-base-dev gfortran
sudo pip install scipy
sudo pip install -U scikit-learn
Data: Places365, available at: http://places2.csail.mit.edu/download.html
if we store the training data under directory ../places365
To train the model:
python train.py
--seed 100
--dataset places365
--dataset-path ../
--checkpoints-path ./checkpoints
--batch-size 64
--epochs 10
--lr 3e-4
--label-smoothing 1
We adapted some utility functions and skeleton code from Nazeri et al. K. Nazeri, E. Ng, and M. Ebrahimi. Image colorization usinggenerative adversarial networks. pages 85–94, 2018.