This project focuses on automatic colorization for grey images.
Different methods were experimented and compared.
Two specific architectures based on CNN and one based on GAN were tested. CNN architectures seem to be more consistent on colorization but GAN model uses more different colors on the result.
Furthermore, recommendation system base on deep ranking model was used to recommend similar images for any given image, and GAN model was retrained using the recommended references.
In general, the retrained GAN model has the shortest training time and best results.