'DeOldify' reviving monochrome photographs over 100 years ago to bright color photographs with deep learning

Many photographs that are historically valuable materials are black-and-white photographs, and it often happens that they do not understand their coloring while telling the situation at the time. " DeOldify " published by Jason Antic on GitHub is a deep learning based model that can color black and white photographs naturally.

GitHub - jantic / DeOldify: A Deep Learning based project for colorizing and restoring old images
https://github.com/jantic/DeOldify

Pictures actually colored by DeOldify are as follows. Everything is beautifully colored, some of which are incredibly natural colors that the original is a black-and-white photograph.

Ballerinas who played " Sleeping Beauty " premiered in St. Petersburg, Russia in 1890.

Pictures taken in Sweden in the 1920s. A woman is relaxing on the sofa.

Native American men wearing fine feathers. It is a picture around 1898.

Landscape under the Eiffel Tower in Paris in the 1880s. The contrast of the blue sky that can escape with the rebar of the Eiffel Tower is superb.

Pictures that men teach geography to five girls in the 1850's.

People who saw television for the first time at Waterloo Station in London in 1936.

You can color even degraded and rough photos. Picture of a samurai taken in 1860

DeOldify can color not only black-and-white photographs but also monochrome line drawings automatically. Below is a drawing of the scenery of the opium cave seen in China in the 19th century.

However, DeOldify is not necessarily perfect. For example, the picture of a woman of Native American who was taken in 1908 seems to be a natural coloring at first sight, but looking closely, the right hand of a woman has been dyed red.

According to Antic, DeOldify is a model based on deep learning, and it is said that Self-Attention type Generative Adverarial Networks (Generative Hostile Network, GAN) is used. In addition, the part published to GitHub this time is still in the early stage of learning, and it seems that we will learn so that we can process not only color but also contrast and brightness as in the following image. Antic said DeOldify said that this color photography is only a part of the data pipeline that Antic is handling and should help in the future to fix every image.