Seems that the letters are all different numbers of pixels, which looks problematic.
This is not problematic in itself, using a (mosified) version of the digitsearch function in BetPot, and a library of all the letters in BMP format.
The main problem is accurately targetting the words, and the mammoth CPU usage scraping for 26*2 (upper and lower) + 10 (digits) texts + some punctuation too.
I do have kicking about a script to do it, but it ain't fast
It can just about keep up with the chat on 4 tables, on my pretty fast system. It uses shortcuts too, matching whole words (like "Dealer: " to skip over a good chunk of pixels.