Google has been working on a new artificial intelligence to compress images which is being taught how to balance quality with minimal file size.
The neural network was built by Google’s research staff from the Tensor Flow machine system, which Google open sourced in November 2015.
6 million images that had been compressed normally were gathered from the internet and shown to the neural network as a series of reference points. The network would then break down the image into 32×32 pixel pieces.
Next, it examined the 100 least efficiently compressed of these image pieces., judging them on which parts were the least compressed using the PNG file format, which notoriously resists compression. These would usually be the more complex sections of the image, meaning that it would have a far easier time working with the simpler sections. Past attempts at this training had only been able to use larger 64×64 chunks. After its intense study, the neural network was able to mix and match different forms of compression on the different areas of an image.
The Google team say that in standard tests, the neural network can outdo standard JPEG compression. However, a major issue to overcome is that the network is basing its decisions predominantly on quality versus space reduction rather than any natural aesthetics.
This means that while some images can look spectacular after being compressed, others turn out…weird. Human eyes can detect subtle abnormalities that machines still struggle with and could tell when the image was reassembled in a jumbled looking manner. If this happened, the network’s compression attempt was classed as a failure. This resulted in some worrying difficulties in gathering reliable data as the test has intrinsically subjective elements to it as to what classifies as, well, weird.
As is parodied in HBO’s Silicon Valley, mastering compression technology is the holy grail of the digital frontier. Though we are making leaps and bounds in increasing storage size, this is a war on two fronts and files have to get smaller.
With images in the millions being created and stored online daily, it is understandable why Google is putting their time and money into this effort. Having an AI craft a unique superior compression method for each image would be a vastly more palatable option than trying to find the right way to compress the next meme you absolutely have to send.