Given a grayscale photograph as input, this demo attacks the problem of hallucinating a plausible color version of the photograph. This problem is clearly underconstrained, so previous approaches have either relied on significant user interaction or resulted in desaturated colorizations. A fully automatic approach has been proposed that produces vibrant and realistic colorizations. The underlying uncertainty of the problem was embraced by posing it as a classification task and use class-rebalancing at training time to increase the diversity of colors in the result. The system is implemented as a feed-forward pass in a CNN at test time and is trained on over a million color images. The algorithm is evaluated using a "colorization Turing test," asking human participants to choose between a generated and ground truth color image. The method used here successfully fools humans on 32% of the trials, significantly higher than other methodology used by the other photo automation tools. Moreover, the colorization can be a powerful pretext task for self-supervised feature learning, acting as a cross-channel encoder. This approach results in state-of-the-art performance on several feature learning benchmarks.
richzhang/colorization is licensed under the
A permissive license that comes in two variants, the BSD 2-Clause and BSD 3-Clause. Both have very minute differences to the MIT license.
[](https://github.com/richzhang/colorization)