Leveraging Transfer Learning to Imitate Human Behavior for Image Colorization
We implemented an Ensemble model for image colorization, the task of generating colored images from grayscale image inputs.
We built upon the work by Melas-Kyriazi and Han 
We leveraged transfer learning for the classification, Alexnet 
By treating image colorization as a classification task and using other advancements, the best results are possible . Regression task settings are easier but perform worse . We tried to imitate human behavior, first classify the scene and then choose sensible colors, with a regression-based (MSE loss) Ensemble model that adds separated classification results as mid-features (Fusion Layer).
For references see the poster.