Scanning a paper document returns about 96% white pixel. Some paper with grayscale DTP or photos will be wildly different.
Old fag goes back to fax tech and the Tel Aviv Tech that tasked mathematician team to code up a solution for security - read graphic (dumb color) file of traffic vehicle plates and render into ASCII for dbase search>identification west bank origin>police check. Well it sounds way way way easier than reality. Anyway that all got sold off for speed cameras and traffic citations around the world…
Back on topic…
Text on paper is a load easier task. Recognise text from random graphic. Allow for font kerning and identification, id validation test points against image library lookup/image retest>image valid>add to validated ASCII output.>apply to commands.
https://en.wikipedia.org/wiki/Comparison_of_optical_character_recognition_software
see https://en.wikipedia.org/wiki/Automatic_number-plate_recognition
anyways OCR Optical Char Recognition shelf is now stacked with Open Source for these Dominion crew.
Again, if you want to vandalise the Trump vote Coomer could mess with ocr process for individual letters (reported calc 250milliseconds per letter) making up Trump or otherwise dump out votes.