A new Google technology allows you to zoom images even if their resolution is small.
No, we are not going to make any jokes about CSI (except the one in the title); We are dealing with a technology that will make the fantasy of creating a discernible image out of just a few pixels a reality. For practical purposes, it allows you to zoom immensely higher than we are used to; to the point where it looks like you’re creating pixels out of thin air.
Google Brain takes full images from pixels
The project (pdf) comes from the hand of Google Brain, and as its name indicates it is based on neural networks; These systems are connected to each other, imitating the structure of the neurons in our brain. It is the next step in creating ultra-powerful systems that achieve what until now was science fiction.
At Google, they took advantage of these neural networks to explore the limit of image enlargement; even when our own brain cannot discern what is in the original image. Imagine an image that is 8 pixels high by 8 wide; The Google Brain system is capable of enlarging it and creating a 32 x 32 pixel image. And best of all, is that the end result closely resembles the actual image.
It is easier to show it than to tell it. Look in the first column, where the original images are reduced to 8 x 8 pixels; in the middle column are the images created by the neural network. The surprising thing comes when we compare them with the original images, in the third column. Ok, in some cases the similarity is not very great, but it is still surprising considering the base from which it starts.
Sure, you might be wondering how the Google Brain system is capable of get details where there are none; The answer is that it actually cheats, because it is not based only on the 64 pixels of the image that we give it.
How Google’s neural networks get to zoom images of a few pixels
In reality the system is made up of two neural networks; the first one is in charge of map the 8 x 8 pixel image with other high resolution images you are looking for on the Internet. Yes, surely you know that Google Images allows you to search using an image, and this neural network does something similar.
The second neural network uses a library called PixelCNN to generate an image based on the 88 image and the high resolution images discovered by the first network; these are part of a database in which the network is fixed as the original image is enlarged. According to find similar images, the system adds pixels from them to improve the original.
In the tests, the results were more than decent. The volunteers were presented with two images of people, one the original and the other the enlarged; the system managed to cheat them 10% of the time (50% would be considered a perfect result). This result was even better when they were presented with two photos of a room; in this case, the system cheated on them 28% of the time. Of course, both times it was far superior to other image magnification systems.
Although this system would be very useful to someone who found a very poor resolution image, it has its negative points; the main one which, as the resulting image has pixels from other images, is not entirely accurate and it would not be worth for the police to enlarge images of suspected suspects. We already told you not to make many CSI jokes.