Neural network trained to independently complement images

An example of the work of a neural network.

English startup Magic Pony Technology claims that it has developed a revolutionary technology of "modeling" images, which significantly improves the resolution of photos and video in real time.

A trained neural network does not just interpolate pixels, but adds the missing details. Developers say that you can automatically generate elements for realistic virtual worlds, for example.

One of the examples of the work of a neural network is given above - the face of the hero of a computer game, originally strongly pixelated.

Another example is shown in a short video . On the left - the original video stream of the game, on the right - improved graphics.



According to the MIT Technology Review, when training a neural network, researchers generated low-quality images from high-resolution images - and submitted these pairs of images to the input of the neural network. As a result, the neural network gradually learned to detect patterns and perform the reverse operation. The system does not require manual marking.


The image on the right is generated on the basis of the sample on the left.

Company co-founder Rob Bishop says that they are now negotiating with several major streaming operators about technology licensing.

In the future, technology can be integrated even on mobile phones, which are also gradually being equipped with fast GPUs. Here, the program can effectively improve the quality of photos taken in poor lighting conditions or with low resolution matrix.

On the application of the generation of landscapes in virtual worlds already mentioned above. Another potential use of a neural network is in image and video compression formats. However, for free formats you will have to create a free analogue, because it uses proprietary technology. The company Magic Pony Technology has filed for more than 20 patents.