An supervised night to day image translation using Deep Convolutional Generative Adversarial Network. This code was built as my final Bachelor Thesis at Telkom University.
- Python 3
- Tensorflow
- CPU or NVIDIA GPU + CUDA CuDNN
git clone https://github.com/evanezcent/Night-to-Day-Image-Translation-using-DCGAN/
pip install tensorflow
pip install PIL
pip install pydotplus
pip install IPython
pip install numpy
pip install pandas
We provided our augmented datatrain and datatest which has a square shape. If you want the un-augmented data, just email me here.
On data preparation, we extract many timelapse video and then tak 5~10 copies of the image.
We just doing two kind of augmentation process, that is cropping and flipping to keep the original feature of the images. Then we renamed it sequentially.
We evaluate the model using SSIM as the accuracy and L2-Norm to calculate the loss.
As the final result, we just get a 40% accuracy using datatest. Based on our analytics, it caused of :
- Lack of paired image data night and day
- Training time
- Suitable architecture for night to day translation case
