Author: Damian Parsons

Google researchers have developed a method using a neural network to compress image files more efficiently than current modern methods such as with JPEG.

AI TensorFlow

Researchers built an AI system using Google’s TensorFlow learning machine, then took 6 million random images from the net which had been compressed using modern day typical methods.

How it’s done

Each image was broken into 32×32 pixels and analyzed the 100 pieces with the least efficient compression methods. The idea here is that the AI could learn by looking at the most complex image areas, which makes compression of less complex sections much easier.

After this process was complete, the AI could then predict what an image would look like after compression and then generate an image.

In short, the AI is figuring out how to dynamically compress various parts of an image using varying methods and levels of compression.

According to Google researchers: “As far as we know, this is the first neural network architecture that is able to outperform JPEG at image compression across most bitrates on the rate-distortion curve on the Kodak dataset images, with and without the aid of entropy coding.”

Is it ready for us to use?

The project is still considered in alpha, thus has a long ways to go before it hits us, but it appears to be a step in the right direction towards saving precious memory card space.