The underlying algorithm has not been published. At its core is a color median filter, a block-L1-minimizing generalization of the well-known scalar median filter. Applying such a filter to an image rounds out corners and destroys fine features, so a separate recovery step is needed. Filtered pixels that differ significantly from their original values are iteratively replaced with the distance-minimizing convex combination of neighboring filtered pixels, digging out corners and other destroyed features without re-introducing noise.
It's basically a little image processing algorithm for restoring rasterized vector art that's been degraded by various forms of abuse (noise, JPEG, print & scan), packaged up in a simple single-page application (sSPA? ;)
There's also a photo mode for de-blocking heavily JPEGed photos.