Everyone has seen this in films or police series: investigators notice an interesting detail on a picture or video, they ask for an enlargement and, as if by magic, the killer’s face clearly reflects in the victim’s glasses… Thanks to technological progress this could become reality: Google’s new artificial intelligence (AI) system allows to improve the pixel resolution of a picture.
Functioning
The new AI system can enhance an eight-pixel square image, increasing the sixteen-fold resolution and effectively restoring lost data.
In this picture, the image on the left is what was presented to the algorithm developed by Google: an image of 8 by 8 pixels. On the right, it’s the original image (32 by 32 pixels). And in the middle, it’s the prediction of Google’s AI.
The system was created by showing innumerable images of faces to feed the AI. The goal is that the machine learns to recognize the main typical facial features.
Scientists have also given the machine many poor-quality images, in order to teach it how to recognize a good quality image, through the comparison of 8 by 8 pixels’ images with 32 by 32 pixels’ images.
Thus, Google’s neural networks can be used to increase the resolution of blurred or pixelated faces. For example, Google’s system can recognise what it’s likely to be a pair of lips and draw the image accordingly.
Test results
The AI has been tested with faces of celebrities as well as with rooms for children.
Although the results are impressive, sometimes it doesn’t have enough information to redraw the face of a person who’s represented in the original picture. Consequently, the algorithm makes mistakes and creates from time to time monstrous faces. However, the program knew every time what it had to recompose: a face or a room.
Towards a wider use?
According to Google, this project was a simple test and the company does not plan to use it yet.
Scientists don’t claim having succeeded in reconstructing the original face: the system creates only a face with a realistic appearance in a resolution of 32 by 32 pixels.
The machine tries to deduce the initial appearance of an image and therefore can generate plausible new details for a human observer. However, the system is based on probabilities and sometimes, it might propose multiple solutions.

A propos de Aline DOUARD