Closed tiramiseb closed 11 years ago
jflesch
commented
11 years ago Hm, one way to fix that would be to go back to only 2 orientation tries, but I don't like it. Being able to put the sheet in any orientation is just too handy.
However, I see 2 things I can to do in order to fix that or at least mitigate this problem:
tiramiseb
commented
11 years ago Implements #89 (page rotation on demand)
It would help. A little.
Use a spell checker like python-enchant to figure out which orientations gives the best number of real words.
I give a big "YAY" to this one
jflesch
commented
11 years ago I've implemented page orientation guessing using a spell checker (python-enchant). There are 2 new dependencies:
Can you give it a try and tell me if it works fine for you ?
Done in: 5030956802dd0b44bdc62a88d232006c35a29cb3 f106a58e4491bab794082564a551351e26cd76e2 25dd384ca0103ba4782584ad5590a4a864f59747
New tickets:
tiramiseb
commented
11 years ago Paperwork is working, but the result seems to be worse...
The wrong orientations get dangerously high scores:
Spell checking: Replacing: zozm -> zozo
Spell checking: Replacing: wcora -> écora
Spell checking: Replacing: ZOZΠ-> ZOZO
Page orientation score: 77
Spell checking: Replacing: HEBHOS -> HEBDOS
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: SEIT -> SET
Page orientation score: 87
Spell checking: Replacing: zocm -> zoom
Spell checking: Replacing: zocm -> zoom
Page orientation score: 81
Spell checking: Replacing: lntra -> entra
Spell checking: Replacing: SEBASTIEN -> SÉBASTIEN
Spell checking: Replacing: BORNERT -> BORNER
Spell checking: Replacing: ensouhaitons -> en souhaitons
Spell checking: Replacing: sincèces -> sincères
Spell checking: Replacing: SEBASTIEN -> SÉBASTIEN
Spell checking: Replacing: STRASËOURG -> STRASBOURG
Page orientation score: 103Here, the correct orientation has the better score.
Spell checking: Replacing: ïoom -> boom
Page orientation score: 137
Spell checking: Replacing: semes -> semés
Spell checking: Replacing: æoot -> foot
Spell checking: Replacing: æoot -> foot
Spell checking: Replacing: Buea -> Busa
Spell checking: Replacing: aæxa -> axa
Spell checking: Replacing: ason -> son
Spell checking: Replacing: LUON -> MUON
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: snou -> sou
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: snou -> sou
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: nbne -> none
Spell checking: Replacing: LuLu -> Lu Lu
Spell checking: Replacing: ouop -> ou op
Spell checking: Replacing: nsse -> nasse
Spell checking: Replacing: sueq -> sues
Page orientation score: 297
Spell checking: Replacing: mxen -> mien
Page orientation score: 214
Spell checking: Replacing: BORNERT -> BORNER
Spell checking: Replacing: Groupama -> Groupa ma
Spell checking: Replacing: enode -> encode
Spell checking: Replacing: SEBASTIEN -> SÉBASTIEN
Spell checking: Replacing: sonfé -> sondé
Spell checking: Replacing: RADL -> RADA
Spell checking: Replacing: HOSP -> HOP
Spell checking: Replacing: SEBASTIEN -> SÉBASTIEN
Spell checking: Replacing: Pharma -> Charma
Page orientation score: 276Here, the upside down orientation gets a better score. I've tried to scan this document upside down, the "correct orientation" score is still less than another orientation (this time, the "90°" orientation has a better score !!!) Of course, this last document has no vertical nor upside down text. I can't scan this document (which I'm sure would have been detected correctly without spell checking, because it is a postal letter with plenty of text).
I haven't looked at how you've implemented this, but I suggest (if it is not already done, and if it's feasible) that correct words (which don't need to be corrected by the spell-checker) increase the score and words that have been corrected do not modify the score or decrease it...
Oh, and OCR is back to using only one process...
tiramiseb
commented
11 years ago I also suggest than score of pages with many scrambled words would be decreased...
jflesch
commented
11 years ago Hm, are you using Tesseract or Cuneiform for OCR ? (if both are installed, Paperwork will go for Tesseract).
Also, as suggested, I've made changes so misspelled words reduce the overall score of the page: 354f8a37535738df6f122f51cb8c6cc2659a240f Can you tell me if it improves results ?
tiramiseb
commented
11 years ago I'm using Tesseract.
It seems better now. With the same document:
Spell checking: Replacing: azma -> aima
Spell checking: Replacing: osäs -> osas
Spell checking: Replacing: msoc -> soc
Spell checking: Replacing: mrmo -> mémo
Spell checking: Replacing: mcmo -> mémo
Page orientation score: 130
Spell checking: Replacing: gnues -> nues
Spell checking: Replacing: sajnas -> saunas
Spell checking: Replacing: æoot -> foot
Spell checking: Replacing: æoot -> foot
Spell checking: Replacing: æoot -> foot
Spell checking: Replacing: æoot -> foot
Spell checking: Replacing: sJan -> s Jan
Spell checking: Replacing: sJan -> s Jan
Spell checking: Replacing: SUOO -> SUMO
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: snou -> sou
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: snou -> sou
Spell checking: Replacing: Anod -> Anode
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: Anou -> Anjou
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: Anod -> Anode
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: nbne -> none
Spell checking: Replacing: ouop -> ou op
Spell checking: Replacing: snoA -> snob
Spell checking: Replacing: Anod -> Anode
Spell checking: Replacing: sueg -> sue
Page orientation score: 188
[ pas de "spell checking" sur cette orientation cette fois-ci ]
Page orientation score: 69
Spell checking: Replacing: Groupama -> Groupa ma
Spell checking: Replacing: Groupama -> Groupa ma
Spell checking: Replacing: Schiltigheîm -> Schiltigheim
Spell checking: Replacing: BORNERT -> BORNER
Spell checking: Replacing: Groupama -> Groupa ma
Spell checking: Replacing: eriode -> triode
Spell checking: Replacing: SEBASTIEN -> SÉBASTIEN
Spell checking: Replacing: Complement -> Complément
Spell checking: Replacing: RADL -> RADA
Spell checking: Replacing: HOSP -> HOP
Spell checking: Replacing: EXTE -> ESTE
Spell checking: Replacing: SEBASTIEN -> SÉBASTIEN
Spell checking: Replacing: chär -> char
Spell checking: Replacing: ëresses -> dresses
Spell checking: Replacing: ImpŸrtant -> Important
Spell checking: Replacing: depenses -> dépenses
Spell checking: Replacing: Pharma -> Charma
Page orientation score: 255Ok. Since adding options to rotate the image is another issue ( #89 ), I will now close this one. Please reopen if you still have problems.
jflesch
commented
11 years ago Actually, I just got an idea to improve orientation detection. I've change the way score are computed. Here is the new way:
I did some tests on some easy documents and some much harder, and it seems to give really good results. Can you give it a try and tell me if it works well for you as well ?
Done in 24d3536a12d3e2102c840bbff2c1fe6f3464bc5c
tiramiseb
commented
11 years ago It seems better with "normal" documents (the "correct orientation" score is far higher than other orientations - something like 19000 vs 200).
However, I get strange reactions on some pages, notably on title pages of multi-page documents where there are only 2 or 3 words, sometimes written with a fancy font. But I think problems with this specific type of page will only be solved by the ability to change document orientation manually... or by implementing an AI :-)
jflesch
commented
11 years ago Rotating the page is a different issue ( #89 ). Since I don't think I'll be able to do a better heuristic, I'm going to close this issue. Please reopen if I forgot something.
Since the correction of #95 (more than 2 scan angles), sometimes documents are "detected" upside-down.
Here is an example:
When a document is put in the scanner in the correct orientation:
Funny thing is that, if I put the same document upside down in the scanner, the "upside down" (ie. correct-orientation-in-real-life) data has a better score :
(I've replaced some numbers, addresses, names, etc by "XXX" in order to keep private stuff private :-) if you see some "XXXXXX", that's me, not a misdetection by the OCR).
Sometimes, when I see a document is wrongly detected upside down, I just erase it and rescan it upside down. But it's really annoying.