Google has come up with a brand-new image compression algorithm it devised, the purpose of which is to reduce file size of JPEGs by 35 percent, without any noticeable or significant dips in image quality.
Google's Guetzli Image Encoder
An even more important aspect is the algorithm's compatibility with a range of browsers, devices, apps, and the JPEG standard. Unlike Google's other attempts in image compression such as WebP and WebM, the encoder in question, called Guetzli, or Swiss German for "Cookie," is versatile and will readily work in the aforementioned platforms.
Google Research's Zurich office helmed the project, and if some are pressed with its confectionery-themed original name, then don't be; Ars Technica's analysis didn't render any relation of Guetzli to cookies, or anything of the sort.
How Image Compression Works
There are several approaches to tweaking JPEG images in terms of quality and size, but what Guetzli specifically targets is the quantization phase of image compression wherein there is more loss in terms of visual quality. In a nutshell, this process aims to reduce huge amounts of hard-to-compress disordered data into easy-to-compress ordered data. In typical JPEG encoding systems, this usually involves switching color gradients to single color blocks altogether or eliminating small details on a picture.
In compression, the tradeoff of keeping file sizes relatively small is sacrificing image quality. The trick is for compression systems to strike a balance between reducing the file size while keeping visual integrity regardless of detail removal. Every lossy encoder, as noted by Ars Technica, takes on it with different approaches, and in the case of Guetzli, Google says it does exactly this form of balance.
How Guetzli Approaches Image Compression
Guetzli, in this case, uses a psychovisual — a fancy term to denote the human vision processing system — model called Butteraugli to determine which colors and details to retain, and which to discard. Basically, the model "approximates colour perception and visual masking in a more thorough and detailed way" than other models.
