A neural network was taught to accurately colorize old movies

Researchers from the Graz University of Technology and the company HS-Art have developed an algorithm primarily

automatic but fully controlleduser of the black and white film restoration process. The technology reduces costs, allows you to accurately reproduce the original colors and recreate the “style” of the old shooting.

The presented technology combinesautomatic tape processing with controlled manual coloring of individual frames. Fully automatic film coloring is possible, the scientists explain, but it has the disadvantage that the colors may be beautiful but not true.


Restoring the colors of a black and white film using a neural network. Video: HS-Art

The algorithm is equipped with a sufficiently large collectiontraining samples to automatically take over the colorization of historical films. The technology requires minimal human intervention: for example, the user can specify colors for a particular frame, after which the AI ​​takes over the coloring of the following frames.

With the help of the developed film algorithms, it is possible torestore very accurately, say the developers. But in some cases this is undesirable: to make films look authentic, sometimes you need to add a small amount of noise corresponding to the graininess of the film. AI can independently generate noise on colored images.

The researchers have published the source code of the algorithm in the public domain, but the full version of the neural network used for video processing is available in the paid application of the company HS-Art.

Read more:

One planet can put an end to life on Earth: how scientists have proven the fragility of the solar system

Treasure found that was hidden during the war almost 1,000 years ago

Notre Dame fire helps scientists unravel the secret of the cathedral's construction