For a very long time now, when it’s time to compress a huge, non-moving, high-resolution image and throw it online, you had very few widely supported choices. Formats like TIFF and RAW are only now starting to pick up support and still have huge file sizes, so the long-suffering JPEG format is still the de facto ruler of the internet’s images. Compressing gigantic images to under a megabyte in most cases with fairly minimal loss of quality, the JPEG format is not without its flaws, which mostly manifest as visual artifacts and other weirdness in compressed pictures. Google, armed with the power of neural networking, is apparently aimed at changing this status quo.
Using neural network tech not entirely dissimilar to the tech behind their “deep dream” image modification experiment, with more than a few new tricks up its sleeve, Google has created a few standards that, given proper conditions, can outperform JPEG by an average of roughly 4-8% in image quality byte for byte. Google presents a few different compression architectures that are a bit different under the hood, but spring from a common root; A neural network is responsible for encoding, decoding, binary conversion, and even an entropy generator to give the machine running the compression algorithms a bit of oomph and allow the formats to tackle bigger and more detailed images, using less system resources than comparable work in JPEG format.
In a full white paper, Google goes into excruciating detail about how it all works, putting their algorithms and methods out there in their entirety for others to base their work around. With the paper up for public use not long after Google announced that the cheapest version of their Google Compute Engine just got cheaper, one can’t help but think that parties interested in image compression for commercial purposes may just start making more use of the platform using these new algorithms. While their algorithms and binarization techniques outlined in the paper seem par for the course enough, their reliance on neural networks and the relative lack of prominence of such systems outside of big data operations serve to make Google’s platform very appealing to a number of demographics, such as web designers and artists, who have never given it a second glance before. Interested parties can, of course, hit up the source link for the full, freely-available white paper.
