3 ms·
If the book is just text and headings, can't you expect OCR to do a good job without any real need of human intervention?
by wodenokoto 11y ago
If the book is just text and headings, can't you expect OCR to do a good job without any real need of human intervention?
- vitovito 11y agoDepends on what you mean by a "good job" and what you're doing with the results. 80-90% recognition isn't good enough if it means when you search for a term it doesn't show up because the OCR saw "rn" and wrote "m", or if you're having it translated, or having it read aloud with a text-to-speech synthesizer. In my tests, we were seeing accuracy problems of 10-15% of lines needing correction, and this is a book that was primarily headers and text. Sometimes this is character-level issues, like I cited above. Sometimes this is dust, debris, shadows or markings being confused with text. You get a little closer by running spelling and context checks against the words, but it's never 100% accurate. And if you aren't looking at the original pages, or you need automated systems to search/parse/translate/etc. the text, you need it to be 100% accurate, which means you need a human editor. OCR isn't a solved problem.
- wodenokoto 11y agoThank you for the detailed response. I honestly thought it was a solved problem (as in better performing than a human) as long as there is just running text.