Tesseract: text2image segmentation fault on macOS ( regression #195?)

[atuyosi]

There seems to be a regression bug with text2image of 3.05 release.
The text2image utility caused segfault, so I couldn' use tesstrain.sh on macOS (10.12.3).

$ text2image --text eng.training_text --outputbase=eng.TimesNewRomanBold.exp0 --font='Times New Roman, Bold' --fonts_dir=/Library/Fonts --tlog_level=3
query weight = 700   selected weight =700
query_desc: 'Times New Roman, Bold' Selected: 'Times New Roman, Bold'
Render string of size 6801
Starting page 0
max_width = 3400, max_height = 4600
len = 6801  buf_len = 6801
Found offset = 6421
Segmentation fault: 11

ref. https://github.com/tesseract-ocr/tesseract/issues/195

for detail:

$ lldb /usr/local/bin/text2image
(lldb) target create "/usr/local/bin/text2image"
Current executable set to '/usr/local/bin/text2image' (x86_64).
(lldb) run --text eng.training_text --outputbase=eng.TimesNewRomanBold.exp0 --font='Times New Roman, Bold' --fonts_dir=/Library/Fonts
Process 40801 launched: '/usr/local/bin/text2image' (x86_64)
Process 40801 stopped
* thread #1: tid = 0xffe117, 0x0000000100bdd5a7 libpangoft2-1.0.0.dylib`pango_fc_font_get_glyph + 25, queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, address=0x0)
    frame #0: 0x0000000100bdd5a7 libpangoft2-1.0.0.dylib`pango_fc_font_get_glyph + 25
libpangoft2-1.0.0.dylib`pango_fc_font_get_glyph:
->  0x100bdd5a7 <+25>: movq   (%rcx), %rdi
    0x100bdd5aa <+28>: testq  %rdi, %rdi
    0x100bdd5ad <+31>: je     0x100bdd5b8               ; <+42>
    0x100bdd5af <+33>: movq   %rax, %rsi
(lldb)

$ tesseract -v
tesseract 3.05.00
 leptonica-1.74.1
  libjpeg 8d : libpng 1.6.28 : libtiff 4.0.7 : zlib 1.2.8

[amitdo]

Here is the source for the regression:
https://github.com/tesseract-ocr/tesseract/commit/709935851061#diff-b37dca9f063c3727f62c496e514177a9L440

[amitdo]

It also affects 4.00.

[amitdo]

@atuyosi, please try my patch:
https://github.com/amitdo/tesseract/commits/3.05-fc-fix
https://github.com/amitdo/tesseract/commit/da7a7fa2f23f

[atuyosi]

Hi @amitdo

Thanks for your quick response.

I tried your patch, but same result.

$ wget https://github.com/amitdo/tesseract/commit/da7a7fa2f23f.patch
$ patch -p1 < da7a7fa2f23f.patch
patching file training/pango_font_info.cpp

try patched binary:

$ text2image --text eng.training.txt --outputbase=eng.TimesNewRomanBold.exp0 --font='Times New Roman, Bold' --fonts_dir=/Library/Fonts --tlog_level=3
query weight = 700   selected weight =700
query_desc: 'Times New Roman, Bold' Selected: 'Times New Roman, Bold'
Render string of size 84
Starting page 0
max_width = 3400, max_height = 4600
len = 84  buf_len = 84
Segmentation fault: 11

$ lldb /usr/local/bin/text2image
(lldb) target create "/usr/local/bin/text2image"
Current executable set to '/usr/local/bin/text2image' (x86_64).
(lldb) run --text eng.training.txt --outputbase=eng.TimesNewRomanBold.exp0 --font='Times New Roman, Bold' --fonts_dir=/Library/Fonts --tlog_level=3
Process 88534 launched: '/usr/local/bin/text2image' (x86_64)
query weight = 700   selected weight =700
query_desc: 'Times New Roman, Bold' Selected: 'Times New Roman, Bold'
Render string of size 84
Starting page 0
max_width = 3400, max_height = 4600
len = 84  buf_len = 84
Process 88534 stopped
* thread #1: tid = 0x122eeb9, 0x0000000100cfb5a7 libpangoft2-1.0.0.dylib`pango_fc_font_get_glyph + 25, queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, address=0x0)
    frame #0: 0x0000000100cfb5a7 libpangoft2-1.0.0.dylib`pango_fc_font_get_glyph + 25
libpangoft2-1.0.0.dylib`pango_fc_font_get_glyph:
->  0x100cfb5a7 <+25>: movq   (%rcx), %rdi
    0x100cfb5aa <+28>: testq  %rdi, %rdi
    0x100cfb5ad <+31>: je     0x100cfb5b8               ; <+42>
    0x100cfb5af <+33>: movq   %rax, %rsi
(lldb)

[amitdo]

Try to manually set the environment variable PANGOCAIRO_BACKEND to fc and then call text2image.

[atuyosi]

No error occurred. It looks work. It seems to be working.

$ PANGOCAIRO_BACKEND=fc text2image --text eng.training.txt --outputbase=eng.TimesNewRomanBold.exp0 --font='Times New Roman, Bold' --fonts_dir=/Library/Fonts --tlog_level=3  2&> detail_log_736_1.txt

detail_log_736_1.txt

But macOS's Preview.app can't open this tif file (although Gimp can open this file ).
I'll try to run tesstrain.sh later.

[atuyosi]

I tried to run tesstrain.sh with above patch and workaround . It works fine. Successfully generated jpn.traineddata.

$ TESSDATA_PREFIX=/usr/local/share/  PANGOCAIRO_BACKEND=fc ./training/tesstrain.sh --lang jpn --langdata_dir ../github_tesseract/langdata/  --training_text ../github_tesseract/langdata/jpn/jpn.training_text --fonts_dir /Library/Fonts/

Note: I replaced mktemp with gmktemp (coreutils ) on macOS

[amitdo]

It should also work without my patch.

The patch was supposed to programmatically set the said environment variable. Since you reported it did not work, I asked you to do it manually. The call to the setenv function in my patch needs to be moved to another place in the code.

[heyidi]

thank you amitdo. it fixed my problem too~~

[FernandoGOT]

Got the same problem on MAC OSX 10.13.3 using tessercat 4.0.0-beta.1, fixed it using the PANGOCAIRO_BACKEND=fc before executing the tesstrain.sh file

[amitdo]

@FernandoGOT,

Can you please check my new patch?
https://github.com/amitdo/tesseract/commit/ee2d6739cf0782

Don't set the environment variable PANGOCAIRO_BACKEND when you test this patch.

[stweil]

Instead of setting PANGOCAIRO_BACKEND=fc, it should also be possible to replace pango_cairo_font_map_get_default() by pango_cairo_font_map_new_for_font_type(CAIRO_FONT_TYPE_FT).

[amitdo]

@stweil, your suggested method looks simpler. Please make a PR.

[zdenop]

Is this issue still valid? If yes can we get this modification to 4.0 finale release?

[amitdo]

I don't have a Mac, but we didn't done anything to fix this issue.

@stweil, https://github.com/tesseract-ocr/tesseract/issues/736#issuecomment-381436973

[atuyosi]

@zdenop It has not fixed (4.0.0-beta.4-164-g5dfce).

It is still necessary to set environment variables PANGOCAIRO_BACKEND=fc .

[zdenop]

Can you please check if 345e5ee1f3e78d16927a667212623d9507cc4a63 solve the problem?

[atuyosi]

@zdenop ,

I tested 345e5ee. No error occurred.

It's OK.

$ text2image --fonts_dir /System/Library/Fonts/  --list_available_fonts
<skip>
$ echo $?
0

$ text2image --text jpn_test.txt  --outputbase=jpn.HiraginoMinchoProLicht.exp0 --font='Hiragino Mincho Pro Light' --fonts_dir=/System/Library/Fonts --tlog_level=3 --ptsize 32
<skip>
$ echo $?
0

Software Version:

$ text2image --version
4.0.0-beta.4-167-g345e

$ sw_vers
ProductName:    Mac OS X
ProductVersion: 10.14
BuildVersion:   18A391

[zdenop]

I have to revert this patch because of side effects. Other solution should be found or implemented...

[zdenop]

@atuyosi: can you please try get the latest code and post output of text2image -v

[zdenop]

I have ready something like this:

diff --git a/src/training/text2image.cpp b/src/training/text2image.cpp
index 1142bb68..3e2b8a23 100644
--- a/src/training/text2image.cpp
+++ b/src/training/text2image.cpp
@@ -50,6 +50,9 @@
 #include "tlog.h"
 #include "unicharset.h"
 #include "util.h"
+#ifdef __APPLE__
+#  include "TargetConditionals.h"
+#endif

 // A number with which to initialize the random number generator.
 const int kRandomSeed = 0x18273645;
@@ -689,5 +692,19 @@ int main(int argc, char** argv) {
     }
   }
   tesseract::ParseCommandLineFlags(argv[0], &argc, &argv, true);
+#ifdef TARGET_OS_MAC
+  // Respect enviroment variable. could be:
+  // fc (fontconfig), win32, and coretext
+  // If not set force fontconfig for Mac OS.
+  // See https://github.com/tesseract-ocr/tesseract/issues/736
+  char* backend;
+  backend = getenv("PANGOCAIRO_BACKEND");
+  if (backend == NULL) {
+    putenv("PANGOCAIRO_BACKEND=fc");
+  } else {
+    printf("Using '%s' as pango cairo backend based on enviroment "
+           "variable.\n", backend);
+  }
+#endif
   return Main();
 }

Maybe we can implement it as option --pangocairo_backend for text2image, so user will be aware about this possibility to change backend.

[amitdo]

    } else {
      printf("Using '%s' as pango cairo backend based on enviroment "
             "variable.\n", backend);
   }

text2image's code needs the 'fc' backend on linux and mac.
Without the 'fc' backend there will be a crash.

[zdenop]

@amitdo : code you quoted means that program will accept what user will set (intentionally). This will give user option to choose backend (in case of new backend or fix for actual)

I have possibility to test it (just sort text - not whole training) on linux and windows:

• On linux (opensuse 15) pango choose CAIRO_FONT_TYPE_FT automatically and I did not recognized any crash. I see no reason to modify code for linux.
• On windows (10) pango choose CAIRO_FONT_TYPE_WIN32 automatically and I did not recognized any crash. Setting it to CAIRO_FONT_TYPE_FT (with PANGOCAIRO_BACKEND=fc). Anyway I prefer to use native solution if it works.

[zdenop]

@tfmorris : Can you please test this ('text2image -v' and patch above?)?

[amitdo]

Linux always uses 'fc'.

Originally Windows had the same issue:
https://github.com/tesseract-ocr/tesseract/issues/380

This patch fixed the crash on windows & mac:
https://github.com/tesseract-ocr/tesseract/commit/8d2d94e4ed32c

but Ray partly reverted it, so it is only compiled for win32 and not on linux or mac.
https://github.com/tesseract-ocr/tesseract/commit/709935851061#diff-b37dca9f063c3727f62c496e514177a9L440

Although the workaround is still there for win32, it has this remark:

// Fixme! Leaks memory and breaks unittests.

[zdenop]

@amitdo: thanks for analysis. Thank I will not limit this change to Mac.

[amitdo]

The new commit will still crash on mac...

[zdenop]

Why?

[amitdo]

Because the default bacend on mac is 'coretext', but text2image is not designed to work with it

You should force 'fc' as backend, at least for mac.

[zdenop]

... and the patch DOES NOT care about default backend... It cares about environment variable definition. So it should work.

[zdenop]

@amitdo: do you have Mac or did you test it on Mac? Because you are claiming that Mac has set PANGOCAIRO_BACKEND variable by default to some value...

[zdenop]

@atuyosi: can you please test current code?

[zdenop]

...and patch, that reflect above logic, will do:

1. if env var is set = do nothing (because this PANGOCAIRO_BACKEND must be set by user and he should know what he is doing)
2. if env var is not set = set it fc (everything should works).

So crash should happened only in case 1. when user set PANGOCAIRO_BACKEND to wrong backend.

[zdenop]

@amitdo : can you please read my comments? You miss the point. Or try the code ;-)

[amitdo]

I will reread the code, maybe I missed something.

[amitdo]

It looks okay :-)

Please pardon me.

You changed the code a few times in the last week, and I was reading the diffs, so I didn't see the whole picture. That was the reason for my confusion.

[stweil]

I have to revert this patch because of side effects. Other solution should be found or implemented...

@zdenop, do you remember which side effects your initial patch had? It fixes pango_font_info_test for MacOS (see discussion for PR #3189).

[amitdo]

See #1999 and #2009

[zdenop]

@stweil : I do not remember, but pango was always strange for me. ;-) After pango changed build system, I lost interest to to play with as I am mainly on windows at this time...

[amitdo]

After pango changed build system, I lost interest to play with as I am mainly on windows at this time...

Meson is great and it does work on windows.

[stweil]

Thank you for your answers. The main problem was obviously unreleased memory. I now have a different solution for #3189.