Image colorization using deep learning techniques is a common real-world application nowadays. In image coloring, a black and white, that is, a grayscale, image is converted into a colored image that best represents the semantic colors of the input image. For example, the color of the sky on a clear sunny day must be colored as blue and not red by the model. There are many colorization algorithms and techniques available; these techniques mostly differ in the way they treat the data and map the grayscale to colors. Some of the parametric methods learn representations by doing training on huge datasets of colored images, posing the problem as regression or classification, and providing proper loss function. Other methods rely on defining one or more color reference images.

In this recipe, we will use autoencoders to achieve this task. We will train the autoencoder with a sufficient number of grayscale photos as input and the corresponding colored pictures as output so that it could discover the hidden structure on correctly applying colors.