Yup, the adaptive thresholder is designed to work with printed pages. That's why I suggested you threshold to bitonal yourself rather than stopping at greyscale.
Tom On Friday, April 22, 2016 at 1:35:12 PM UTC-4, Jonas Pfannschmidt wrote: > > So I did a few more tests and I found out that I can improve the results > by adding a black box (see attached image). If I combine my findings and > Toms answer I come to the conclusion that tesseract probably chooses the > wrong threshold because the image has not enough contrast > > Thanks again!. > > On Friday, 22 April 2016 12:14:36 UTC+1, Jonas Pfannschmidt wrote: > >> Thanks Tom. I tried your suggestion and it does work better. Python code >> for converting the image is here if someone is interested: >> https://github.com/JonasPf/ocr_testtool/blob/master/captest/ocr.py >> >> While the problem is solved, I would still like to understand it a bit >> better. In 'fails.png' it doesn't even recognize the first row (Timesheet, >> Categories, ...) but in 'works.png' it does. Even though this part is the >> same in both images. I tried to cut out only that part from 'fails.png' and >> surprisingly the text gets recognized! So something in 'fails.png' throws >> it off completely to the point where it doesn't recognize text that it >> normally would if that something wasn't there. Any idea what that something >> is? >> >> On Wednesday, 20 April 2016 17:59:35 UTC+1, Tom Morris wrote: >>> >>> On Wednesday, April 20, 2016 at 2:42:58 AM UTC-4, Jonas Pfannschmidt >>> wrote: >>>> >>>> Hi, >>>> >>>> I'm trying to automate UI tests using OCR. The goal is to have a test >>>> script with lines like: "click_text('Reports')" and it automatically >>>> clicks >>>> on the button "Report". >>>> >>>> It works quite well ... sometimes. I've attached two sample screen >>>> captures. The text on 'works.png' gets recognized reasonably well, >>>> 'fails.png' returns only garbage. Both images have been created >>>> programmatically in the same way (capture screen, resize by factor 4, >>>> convert to greyscale). Does anybody know why one works and the other >>>> doesn't? >>>> >>> >>> Well, for one thing, the image that works has a lot more dark text on >>> it, whereas the one that doesn't not only has less text, but some of the >>> text that it has is greyed out. >>> >>> At the end of the day Tesseract is going to be working on a bitonal >>> image, since you've got a non-traditional application, I'd think you'd want >>> to control as much of the image preprocessing as possible to make sure it's >>> getting done in a way that's appropriate for your application, so rather >>> than converting to greyscale, you should threshold and convert all the way >>> down to bitonal. >>> >>> Tom >>> >> -- You received this message because you are subscribed to the Google Groups "tesseract-ocr" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at https://groups.google.com/group/tesseract-ocr. To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/bfe956a8-ee60-4585-99f9-524e79caf276%40googlegroups.com. For more options, visit https://groups.google.com/d/optout.
