Neural networks have learned to paint black and white photos.
One of the most interesting areas to study in IT is neural networks. Their capabilities are amazing. Not so long ago, Google showed psychedelic images created by DeepDream AI based on real photographs. Some of the generated images are just scary.
One of these images ( you can try to make your own here )
But neural networks can work in a more relaxed mode. For example, to make color black and white shots. In order to teach the system to do this, researchers teach it by displaying a large number of different photos. The neural network selects those objects from color images that are similar to the objects in black and white images, and colors the latter. Now such a project is being developed by researchers from the University of California at Berkeley.
')
According to the results of the project, an article entitled Colorful Image Colorization (PDF) was published. Not all “painted” pictures can be called successful, but some are just perfect. The algorithm of the system developed by the researchers works with a number of common rules (for example, the sky is blue, the grass is green, etc.), plus these algorithms work through many similar photos to get information about the exact color of the desired objects.
Color shots were shown to volunteers who tried to guess which shots were originally black and white, and then they were made in color, and which ones were originally in color. About 20% of the colorized images of the subjects were considered initially color. Interestingly, the network is difficult to paint, for example, a green vegetable on a white plate, or leave a snow-white heron of a certain type.
The sources of the algorithm, documentation and other developments can be found on Github .
Now neural networks do not just color photos, but also generate people's faces and interiors . So, a team of researchers from Indico and Facebook created a neural network capable of “inventing” realistic images. By the way, about the evolution of neural networks can be found here .
One of these images ( you can try to make your own here )
But neural networks can work in a more relaxed mode. For example, to make color black and white shots. In order to teach the system to do this, researchers teach it by displaying a large number of different photos. The neural network selects those objects from color images that are similar to the objects in black and white images, and colors the latter. Now such a project is being developed by researchers from the University of California at Berkeley.
')
According to the results of the project, an article entitled Colorful Image Colorization (PDF) was published. Not all “painted” pictures can be called successful, but some are just perfect. The algorithm of the system developed by the researchers works with a number of common rules (for example, the sky is blue, the grass is green, etc.), plus these algorithms work through many similar photos to get information about the exact color of the desired objects.
Color shots were shown to volunteers who tried to guess which shots were originally black and white, and then they were made in color, and which ones were originally in color. About 20% of the colorized images of the subjects were considered initially color. Interestingly, the network is difficult to paint, for example, a green vegetable on a white plate, or leave a snow-white heron of a certain type.
The sources of the algorithm, documentation and other developments can be found on Github .
Now neural networks do not just color photos, but also generate people's faces and interiors . So, a team of researchers from Indico and Facebook created a neural network capable of “inventing” realistic images. By the way, about the evolution of neural networks can be found here .
Source: https://habr.com/ru/post/392619/
All Articles