Latin.traineddata(best) - Words missing in OCR

[TheSeiko]

Environment

• Tesseract Version: tesseract-ocr-setup-4.0.0-alpha.20170804.exe - Latin.traineddata from best
• Platform: Win10 64bit

Current Behavior:

Tesseract skippes words when doing OCR

OCR-Result: USA

Expected Behavior:

OCR-Result: USA/NORDKOREA

Suggested Fix:

not to skip words

---

Today I saw multiple times, that tesseract skips words, sometimes in the middle of a paragraph.

i.e. tesseract test/1502433849760_1.png test/1502433849760 -l Latin

1502433849760.txt

[amitdo]

Known problem for all langs/scripts
#681

[amitdo]

In your case the problem is probably in the layout analysis stage.

[TheSeiko]

I've another example:

tesseract test/1502442621178.png test/1502442621178 -l Latin

Here a word in the middle of the sentence is skipped.

1502442621178.txt

[amitdo]

Try cutting the non-text areas with gimp and retest.

[TheSeiko]

Only the red areas are processed by tesseract, they are written to separate png-files. (I've uploaded the separate files as well and the resulting output-files)

[TheSeiko]

The interesting thing is, two similar frames were processed without problems:

[amitdo]

Here a word in the middle of the sentence is skipped.

Which word?

Might be related to #681 in this case.

[TheSeiko]

Pöllan - the ö exists in unicharset

Thank you for the link to #681

[amitdo]

Try to debug with stopper_debug_level=2

https://github.com/tesseract-ocr/tesseract/blob/3ec11bd37a56/ccmain/linerec.cpp#L293

[TheSeiko]

Best choice certainty=-2.96366, space=-0.206939, scaled=-20.7456, final=-20.7456
: Pöllan : R=11.9817, C=-2.96366, F=1, Perm=2, xht=[0,3.40282e+038], ambig=0
pos NORM NORM NORM NORM NORM NORM
str P ö l l a n
state: 1 1 1 1 1 1
C -0.237 -2.964 -0.232 -0.209 -0.192 -0.199
Deleting word with certainty -20.7456
: Pöllan : R=11.9817, C=-20.7456, F=1, Perm=2, xht=[0,3.40282e+038], ambig=0
pos NORM NORM NORM NORM NORM NORM
str P ö l l a n
state: 1 1 1 1 1 1
C -0.237 -2.964 -0.232 -0.209 -0.192 -0.199

test.txt

[TheSeiko]

Best choice certainty=-0.106016, space=-2.97242, scaled=-20.8069, final=-20.8069
: USA : R=2.24336, C=-0.106016, F=1, Perm=8, xht=[0,3.40282e+038], ambig=0
pos NORM NORM NORM
str U S A
state: 1 1 1
C -0.090 -0.086 -0.106
Best choice certainty=-0.229416, space=-2.97242, scaled=-20.8069, final=-20.8069
: /NORDKOREA : R=16.2872, C=-0.229416, F=1, Perm=2, xht=[0,3.40282e+038], ambig=0
pos NORM NORM NORM NORM NORM NORM NORM NORM NORM NORM
str / N O R D K O R E A
state: 1 1 1 1 1 1 1 1 1 1
C -0.223 -0.202 -0.207 -0.201 -0.192 -0.208 -0.197 -0.193 -0.229 -0.228
Deleting word with certainty -20.8069
: /NORDKOREA : R=16.2872, C=-20.8069, F=1, Perm=2, xht=[0,3.40282e+038], ambig=0
pos NORM NORM NORM NORM NORM NORM NORM NORM NORM NORM
str / N O R D K O R E A
state: 1 1 1 1 1 1 1 1 1 1
C -0.223 -0.202 -0.207 -0.201 -0.192 -0.208 -0.197 -0.193 -0.229 -0.228

output_1502433849760_1.txt

[amitdo]

So it does recognize them, but still decides to drop them...

[amitdo]

'Pöllan' is dropped because it's not in the dictionary and the 'ö' has low certainty.
'/NORDKOREA' is dropped because it's not in the dictionary and has low space certainty.

[amitdo]

'USA' shares the same low space certainty with '/NORDKOREA' but escapes from punishment because it's in the dictionary.

[Shreeshrii]

@TheSeiko Please try with the changes suggested in #681 (comment) to see if you get improved recognition of these words without impacting others.

[amitdo]

After applying my patch:
'/NORDKOREA' ('USA /NORDKOREA')
and
'Pöllan'

are recognized in the final text output.

All other words are recognized the same as before.

[TheSeiko]

Thank you to both of you, your help is much appreciated! I'm on holiday till end of next week then I'll try to compile a windows version with the changes you suggested and test it.

[amitdo]

I'll try to compile a windows version with the changes you suggested and test it.

Did you try my suggestion?

[TheSeiko]

I'm still on it.
It takes me some time as I'm used to Java but I only used C/C++/... at University, which is quite some time ago. Then some other projects took my time.
I've already installed Visual Studio 2017 and the Git client.
I'll keep you updated.

[TheSeiko]

I've been able to compile it now and starting a test run against 50k frames.

[TheSeiko]

looks good

[TheSeiko]

Tried it as well with 64bit but there I get some errors but I don't think the problem is the fix:

E:\Tesseract-OCR4.0ab1>tesseract test/1502442621178.png stdout --oem 1 -l Latin
Warning. Invalid resolution 0 dpi. Using 70 instead.
Estimating resolution as 726
Error in pixCreateHeader: height must be > 0
Error in pixCreateNoInit: pixd not made
Error in pixCreate: pixd not made
Error in pixClipRectangle: pixd not made
Error in pixCreateHeader: height must be > 0
Error in pixCreateNoInit: pixd not made
Error in pixCreate: pixd not made
Error in pixClipRectangle: pixd not made
Error in pixCreateHeader: height must be > 0
Error in pixCreateNoInit: pixd not made
Error in pixCreate: pixd not made
Error in pixClipRectangle: pixd not made
Error in pixCreateHeader: height must be > 0
Error in pixCreateNoInit: pixd not made
Error in pixCreate: pixd not made
Error in pixClipRectangle: pixd not made
Error in pixCreateHeader: height must be > 0
Error in pixCreateNoInit: pixd not made
Error in pixCreate: pixd not made
Error in pixClipRectangle: pixd not made
Error in pixCreateHeader: height must be > 0
Error in pixCreateNoInit: pixd not made
Error in pixCreate: pixd not made
Error in pixClipRectangle: pixd not made
Error in pixCreateHeader: height must be > 0
Error in pixCreateNoInit: pixd not made
Error in pixCreate: pixd not made
Error in pixClipRectangle: pixd not made

---

x86 output:
E:\Tesseract-OCR4.0ab1>tesseract test/1502442621178.png stdout --oem 1 -l Latin
Warning. Invalid resolution 0 dpi. Using 70 instead.
Estimating resolution as 726
PATERNION

21-Jähriger geriet mit seinem
Auto in Pöllan zu weit in die
Fahrbahnmitte. Er rammte
das Auto eines 62-Jährigen.

Der Beifahrer des 21-Jährigen
wurde schwer verletzt.

---

I'll try to find the error and keep you updated.

[amitdo]

That one is related to image processing. Seems like a bug on (Windows?) 64 bit environment.

Please open a new issue for that.

[Shreeshrii]

@zdenop Please close this issue.

Words in https://user-images.githubusercontent.com/30631253/29272525-b9762536-8100-11e7-8d82-0961dd49663b.png mentioned above in #1080 (comment) are not being dropped after patch from @amitdo was merged.

PATERNION

21-Jähriger geriet mit seinem
Auto in Pöllan zu weit in die
Fahrbahnmitte. Er rammte
das Auto eines 62-Jährigen.

Der Beifahrer des 21-Jährigen
wurde schwer verletzt.

[amitdo]

Fixed in #1264.