I have a python script. It is being fed a steam of single-letter images.
They are always black-on-white, always vertically oriented, and always square.
They are .jpeg with very, very bad quality, sometimes blurry, sometimes with artifacts. They are not always same dimensions, and letters are not always at the center.
But only letters, not numbers or symbols.
I am currently recognizing them with pytesseract... which is an overkill and is WAY TOO SLOW (i.e. it takes up to 10 seconds to recognize 20-letter series). What I really need here is to spend not more than a couple seconds on 100-letter series, tops.
So the question is -- Is there a faster solution? "re-train tesseract" is not acceptable, since it is needed on that machine to process more complex images in another processes.
Any idea is welcome (I'm not shure about "write your own recognizer", I'm not that professional... but if someone is willing to teach me how...)