OCR of free hand photo of book

[Borneq]

First I test tesseract on file generated as flat image.

I generate Lorem Ipsum text:

5 paragraphs, 452 words 2978 bytes, 24 lines + 4 blank lines, maximal line len in my editor was 135 chars.

Result: 100% accurate but two full stop marks, fantastic.

Next, I rotate image. Only 0.7 degree caused a lot of confusion and minor rotation 0.1-0.6 degree - treat some m as n.

In my book photo images are often rotate up to 3.5 degree.

Worse, text is transformed into curve lines of text like F-distribution

("What function looks like the edge of a paper book sideways? on math.stackexchange.com)

how to work with real photos of books, it is possible as option or thing that is missing in tesseract ?

[Zdenko Podobny]

Tesseract is OCR engine and the user is responsible for preprocessing  - see the documentation.

IMO there is already app (using tesseract) for what you try to do: Text Fairy [1]

[1] https://play.google.com/store/apps/details?id=com.renard.ocr&hl=en

Zdenko

st 31. 1. 2024 o 2:00 Borneq <borucki...@gmail.com> napísal(a):

--
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 tesseract-oc...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/9ac3343e-df3c-432e-8066-af21a20eda1cn%40googlegroups.com.

[Borneq]

I have Linux and prefer batch. I found https://gist.github.com/endolith/334196bac1cac45a4893 (from https://tesseract-ocr.github.io/tessdoc/ImproveQuality.html#examples). It correct recognizes 1.00 degree. How it combine with tesseract?

[Borneq]

I found mzucker/page_dewarp on github - tool for dewarp books and convert color to black and white