Help with auto island detection

[Mattia Mirri]

Hi All,

I'm coding with C#

im trying to use tesseract to extract various comment from image like this one: 

Here an example of output.. as you can see i expect 6 different rows (one row for each island of text)

Can anyone help me to find the documentation to implent a correct detection of various islands?

Thanks

[Ger Hobbelt]

Relevant (if only sideways at first glance):

- https://tesseract-ocr.github.io/tessdoc/ImproveQuality.html 🎯

- https://groups.google.com/g/tesseract-ocr/c/B2-EVXPLovQ/m/lP0zQVApAAAJ

as the latter article says: tesseract has been engineered for plain text books processing. That's a whole lotta assumptions all rolled into one!

You are very probably running head-on into the tesseract's first stage: page segmentation (bounding boxes production), which is text line oriented and is not plenty sophisticated enough to recognize multicolumn glossy magazine-style layouts (which your page looks like, more or less ;-) ) -- I'm living in hopes of someone taking on that section of the tesseract codebase and doing a real number on it. But one can dream ...

Ergo: preprocessing ftw! Have a preprocess stage extract each column of text into individual images and feed each to tesseract. That should at least get rid of the "jumping columns" behaviour you see in your output.

(Also *postprocessing*: use HOCR or TSV and not TXT output: everybody blogs great & happy news using the latter because simple sells; it is not simple. And when you have HOCR or TSV output, you get glyph/word coordinates and that *can* be a hint: when the 'spacing' is 'weird' (large), then the next word in the output is probably not actually part of the same sentence/column. 

However, it's much much better to prevent than to "fix it in post", so cut that page into separate column images and that's one less problem area to worry about.

Second, there's the hyphenation done by the printer/publisher to get a slightly more palatable column layout for these über-narrow columns: tesseract does not know/care about this trick, so that's for the postprocessing to take care of. Yet another reason for HOCR/TSV vs. TXT: there's oftentimes a difference between hyphenated words and interjection/commentary dashes and their relative placement vs. the previous word-particle: it's a low probability thing but relative spacing sometimes does provide a hint which it is; classical style guides mention em and en dashes but lazy/less sophisticated print, particularly from the 70's and 80's, uses the same bloody hyphen dash for everything.  Ah, just my opinion as a classical quality printing aficionado (I wish I was that good at layout!) and OCR obviously has a problem keeping those three apart anyhow.

... it's 08:10 and clearly time for my medication (*nuzzles the neck of an exquisite whiskey bottle*); gotta go.

Ciao and have fun!

Met vriendelijke groeten / Best regards,

Ger Hobbelt

--------------------------------------------------
web:    http://www.hobbelt.com/
        http://www.hebbut.net/
mail:   g...@hobbelt.com
mobile: +31-6-11 120 978
--------------------------------------------------

--
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 visit https://groups.google.com/d/msgid/tesseract-ocr/754d9a2b-9a24-429b-8292-04f4d15560a2n%40googlegroups.com.