Google Brain, new Google success in field of artificial intelligence. SAN FRANCISCO. See how Google Brain technology woks. February 8 morning news, Google has a news from its artificial intelligence department. “Google Brain” designed a new software, you can restore the same mosaic source picture into a clear picture. In short, the original “code” of the photos can now easily become a code-free HD. This uses Pixel Recursive Super Resolution.
First look at a picture, the left column is a resolution of 8 × 8 coding picture, the middle one is the Google Brain software to restore the picture. And the right one is the real original picture. The actual effect is very close to the original image.
As we all know, we can not produce more detailed picture than the source picture. Here the Google brain is how to do it? They intelligently integrate the two neural networks together.
The first part is the conditioning network, which compares an 8×8 image to a high-definition image, shrinking the other HD pictures to an 8×8 resolution, and then matching.
Google Brain Working Principle?
The second part is the prior network, which uses PixelCNN (pixel-based neural network) to add real HD detail to an 8×8 source image. In this case, the priority network is actually absorbing a lot of celebrity and bedroom photos. Then, when the source image needs to be parsed, it will look for new pixels matching it from its known picture and add it.
For example, the image has a brown pixel at the top, and the priority network might think it is an eyebrow: which then causes the image to be resolved. It will have filled with eyebrow-shaped brown pixels.
In order to produce the final code-free high-definition images, the need to integrate the output of these two neural network data. The end result will often contain some specious new details.
Google’s high-definition reduction technology of the brain has been in the actual test have achieved some success. When the team presented real HD star photos and computer-restored photos to human observers. Observers were tricked to 10 percent (50 percent of them were perfect scores). In the bedroom photos, the human observer cheated the proportion of 28%. Both scores are much higher than the conventional interpolation technique, which does not deceive any human observer.
Not actual Pictures Restored by Google Brain
Using these algorithms in the computer restore the high-definition picture is not a true picture. It adds details by only a “guess.” This raises a number of interesting questions, particularly in the area of surveillance and forensics. This technology can change the blur suspect photos into a more clear picture. But this can not get the real suspect photos, but still can provide some help for the police.
Google Brain and DeepMind are Alphabet’s two deep learning institutions. The former recently published some interesting research results. For example, develops a self-made encryption algorithm artificial intelligence technology. The latter is a fame because of its development AlphaGo last year to defeat the world’s top Go players.