Researchers from University of California, Berkeley developed an interactive deep learning-based app that makes it easy to accurately colorize a black and white image in minutes.
Building on the researcher’s previous work of a convolutional neural network automatically adding color to black and white photos, their new app uses the same process, but with the addition of user-guided clues and hints to produce more realistic results.
Using CUDA, a TITAN X GPU and cuDNN with the Caffe deep learning framework, they trained their models on 1.3 million color photos that were made grayscale “synthetically” (by removing the color components).
The app first takes its best attempt to automatically colorize the source image and then creates a small palette of suggested colors. The user is then able to refine the colorization by adding color markers to the source image.
If you want to try it yourself, the researchers published the ‘Interactive Deep Colorization’ code on GitHub. They have also published a variety of historical grayscale photographs that have been colorized with their app and they look incredibly realistic.
Read more >