Arun Pillai

I test on bad scans deliberately, because an engine that reads a clean flatbed page tells you nothing about whether it will help you. Scanning quality decides the ceiling. Skew, uneven lighting, a page photographed at an angle and noise from a cheap scanner all reduce accuracy, and no engine recovers what never made it onto the page. I put that first on every page, because it is the fix that has not happened yet and it costs nothing. The choice between an OCR engine and text detection inside a PDF matters. A PDF made from an image has no text layer, so extraction returns nothing and you need recognition. A PDF made from a born-digital source has a text layer and needs no recognition, and running recognition over it can make it worse. Language selection is not cosmetic. Choosing the wrong language model produces confident, plausible and wrong output, which is worse than failing, because nothing marks it as suspect. Layout analysis is where table extraction loses. Recognition recovers the characters accurately and then hands you a flat stream, so reconstructing a table means inferring the grid from positions. Handwriting is a different problem again, with a much lower ceiling and no tool that solves the general case. Every result here is searchable afterwards, which is the reason for doing it at all. I also cover what to do with a document you should not upload to a hosted service, since OCR usually means sending the file somewhere. Language models for recognition are selected per language and some are far better than others, so a multi-language document needs the right model per page rather than one for the whole file. Getting this wrong produces output that looks like recognition and is not.