Modern Greek data issues
chopinesque opened this issue · 10 comments
There are 2 major issues with the Greek data.
They tend to produce µ (micro sign) instead of μ (Greek m letter) and despite choosing Modern Greek (ell), some characters have accents that belong to polytonic Greek.
https://github.com/tesseract-ocr/langdata_lstm/tree/main/ell contains training text and a word list with the same issues, so the model was trained to produce such results.
Right, so how can this be fixed? For example I can see in https://github.com/tesseract-ocr/langdata_lstm/blob/main/ell/desired_characters and https://github.com/tesseract-ocr/langdata_lstm/blob/main/ell/ell.unicharset the existence of polytonic characters which should not be there.
In a first step you could send a pull request for langdata_lstm which fixes the files there. But finally new trainings are required, maybe based on the existing models for Greek.
OK, I may need some guidance please. I created a fork. So do I simply have to remove non-valid characters from above mentioned files?
I also see
tessedit_load_sublangs grc
https://github.com/chopinesque/langdata_lstm_modern_greek/blob/main/ell/ell.config#L2
I am not sure whether this line should be there going forward.
So do I simply have to remove non-valid characters from above mentioned files?
Remove or replace, what fits better.
Thanks. If I replace, I need to know about the structure, for example,
ὶ 3 0,255,0,255,0,0,0,0,0,0 Greek 124 0 124 ὶ # ὶ [1f76 ]a
How is the 124 0 124 derived?
You can keep the unicharset file unmodified. A replacement will be created when a new training is run.
tessedit_load_sublangs grc
That line tells Tesseract to always use grc in addition to ell. Therefore wrong glyphs can also come from grc as long as that configuration is there.
You can keep the unicharset file unmodified. A replacement will be created when a new training is run.
Not sure then which files I should change. I don't think I have the knowledge to do any training (I also use Windows).
tessedit_load_sublangs grc
That line tells Tesseract to always use
grcin addition toell. Therefore wrong glyphs can also come fromgrcas long as that configuration is there.
So this line should be removed.