Not to detract from the work he did (it's still not easy, and it's good work), but a lot of that could be programmatic---select all pixels that are near-black to extract the text, copy them to a separate layer, enlarge selection by 1.5 pixels, flood fill white, and then paste the original text on top of that. A certain amount of massaging is necessary, but it probably wouldn't be "hours" just for that.

Also, apparently Google has an API that lets you extract specific layers, e.g. just the text (he mentions using this in the article). Makes sense in retrospect, but it's cool to know anyway.