Image compression is an essential piece of the puzzle for mediums that create large files, including 4K video and 360. Functions from live-streaming to web download times are largely dependent on file sizes, but shrinking files involves a balance between quality and space.
Google is teaching computers to recognize the best method for compressing image files, and while the system isn’t (yet) able to judge the effect of a compression on the image’s quality, it is able to analyze what methods are the most successful in adjusting the image’s footprint. The group used 6 million random compressed photos and cropped them down to a 32 by 32-pixel chunk. The neural network then selected 100 photos with the least effective compression, learning what not to do by way of example.
The term “effective” for the AI system may not be exactly the same for photographers looking to shrink file sizes, however — the computer looked at files that, despite compression, still remained quite large. Using the 100 least effective photo chunks, the researchers estimated that the system could learn how to compress simple images.
With that dataset, the system was then able to predict how an image would look after compression, and was further able to create that compressed image itself. Unlike other methods, this approach by Google splits the image into pieces and treats each section independently, selecting the best compression based on how detailed that area of the image is. By compressing the image in sections, the system was able to achieve a smaller file size than traditional compression methods.
While the work looks promising, the size of the file is not the only relevant metric. Measuring the success of each compression based on how the image appears is much less straightforward, since image quality is subjective. The AI system doesn’t look like it’s going to be implemented into something like Google’s cloud storage any time soon, but the study, published last week, does indicate that “smart” compression, where the computer chooses the best method, isn’t too far-fetched.
