Abstract
This paper tackles the challenge of colorizing grayscale images. We take adeep convolutional neural network approach, and choose to take the angle ofclassification, working on a finite set of possible colors. Similarly to arecent paper, we implement a loss and a prediction function that favorrealistic, colorful images rather than "true" ones. We show that a rather lightweight architecture inspired by the U-Net, andtrained on a reasonable amount of pictures of landscapes, achieves satisfactoryresults on this specific subset of pictures. We show that data augmentationsignificantly improves the performance and robustness of the model, and providevisual analysis of the prediction confidence. We show an application of our model, extending the task to videocolorization. We suggest a way to smooth color predictions across frames,without the need to train a recurrent network designed for sequential inputs.