Image colorization is one of the leading-edge topics from the deep learning world. As our understanding of transfer learning and deep learning is maturing, the application scope is getting exciting and more creative. Image colorization is an active area of research and lately some exciting work has been shared by deep learning experts.

In this chapter, we learned about color theory, different color models, and color spaces. This understanding helped us reformulate the problem statement to that of mapping from a single-channel grayscale image to a two-channel output. We then worked toward building a colornet based on the works of Baldassarre and his co-authors. The implementation involved a unique three-layer network consisting of an encoder, a decoder, and a fusion layer. The fusion layer allowed us to utilize transfer learning by concatenating VGG16 embeddings with the encoder output. The network required a few specific preprocessing and postprocessing steps to train on the given set of images. Our training and test datasets consisted of a subset of ImageNet samples. We trained our colornet for a few hundred epochs. Finally, we presented a few hallucinated images to understand how well the model has learned the task of coloring. The trained colornet learned certain high-level objects, such as grass, but did not perform well on smaller or more infrequent objects. We also discussed a few challenges this type of network presents.

This concludes the final chapter in a series of use-case-driven chapters in this book. We presented different use cases across different domains. Each of the use cases helped us leverage the concepts of transfer learning, which was discussed in detail in the first two sections of the book. Andrew Ng, one of the leading names in the field of machine learning and deep learning, said during his NIPS 2016 tutorial that:

Transfer learning would be the next driver for the commercial success of machine learning.

With the variety of applications and their advantages discussed and showcased throughout this book, you should now understand the immense potential of transfer learning.