In movies we are usually used to that cinematic moment in which the protagonist takes a photo camera or a high-profile video camera to start zooming in, cleaning the pixels, and creating a very real image. This was merely pure fiction, although from the Google laboratories it seems that it will be a reality.
The Google Brain lab seems to have come up with a solution that is very close to fiction. The action «Zoom and enhance», means that take a pixelated and noise-filled image, or from a cut-out section of one, by running it through science magic technology, can produce a recognizable face or license plate number.
There is a basic principle in computer graphics that does not allow to extract more data from a higher resolution image, from a very low resolution image. Although this can be erased from our minds thanks to the arrival of machine learning or computational learning.
The task is carried carried out by two neural networks working on the same task from different angles. One network, referred to as "the conditioning", attempts to compare the small 8 x 8 low resolution images to the higher resolution ones by increasing the latter to 8 x 8.
The other network, on the other hand, study specific sets of images, such as the faces of celebrities or rooms, and try to discern patterns, such as the location of the main features of a face. He then applies them to a scaled version of the 8 x 8 image. The two networks are then scaled to a single, higher resolution image, the accuracy of which is quite surprising and most of the time even ridiculous.
The results are quite promising and Google Brain researchers say they are beyond excited. The technique is not completely perfected, but it is a major advance over previous methods.