Surreal is a Deep Learning based super resolution algorithm
The goal for this project is to provide a generic super resolution algorithm for upscaling images. The first implementation is going to use a conditional DCGAN (Deep Convolutional Generative Adversarial Network) so that the generative model operates on perceptual loss - not some kind of hacky euclidean distance metric.
I am reading the papers below but simplifying the model fairly drastically.
Todos
- Adapt super-resolution code for an RGB dataset
- Easy to use data-generator API which is dataset independent.
Done
- Modularize and make the model definition code more object-oriented to make future modifications easier.
- Get DCGAN working on MNIST
- Get super-resolution working for MNIST
References
- Radford, A., Metz, L. and Chintala, S., 2015. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434.
- Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z. and Shi, W., 2016. Photo-realistic single image super-resolution using a generative adversarial network. arXiv preprint arXiv:1609.04802.