This development could result in improved web browsing experience.
Researchers at Google are blurring the lines between fiction and reality: a small team of researchers at Google have built a neural network whose sole purpose is to optimize file sizes in images. They have developed a program that uses artificial intelligence to improve the compression of the JPEG file format so that the file’s overall size undergoes a significant reduction.
Their compression is in fact stretching the very limit of the JPEG file format. Its approach relies on forcing the neural network to learn compression the hard way.
Case in point: Having sampled 6 million compressed photos from the internet, Google’s researchers broke them down into 32 x 32 pixel sizes. The neural network was then fed 100 bits from each image that represented the poorest elements of its compression. This was done keeping in view the hypothesis that if the network could do a better job compressing the worst of the samples, it should be able to do a better job compressing the rest.
In their paper, the researchers have broken down the process using highly complex math which demonstrates how the neural network broke down the images into binary code and then went on to reconstruct them piece by piece, which resulted in the neural network outperforming the JPEG compression at most bitrates.
There is a catch however. Turns out human perception is a bit lossy. As Google says, the “human visual system is more sensitive to certain types of distortions than others,” and moreover there isn’t a universally recognized metric for measuring human perception of a compressed image.
This significant achievement by Google’s researchers could make media in web pages load much faster and that, in my books, is definitely a good thing.