Setting the page segmentation mode to 6 ("Assume a single uniform block
of text") typically improves the layout detection for such texts, but
should not be done in the config file.
unlvtests/runtestset.sh adds `--psm 6` explicitly, so test results
won't change when using that script.
This is similar to commit ecfee53bac.
Signed-off-by: Stefan Weil <sw@weilnetz.de>
While orientation and script detection (OSD) normally requires
osd.traineddata to detect both, it must also be possible to do
only orientation detection with eng.traineddata or any other
traineddata.
Enforce osd.traineddata only if there was no `-l` command line option.
Commit 27ce472666 was too restrictive.
Signed-off-by: Stefan Weil <sw@weilnetz.de>
Setting the page segmentation mode in those config files gives unexpected
results: the text recognized when no config or only txt is given changes
if both txt and any of hocr, pdf or tsv is chosen.
In a test set of nearly 200 pages from historical books, using
segmentation mode 1 is typically slightly better than the default,
but there are also cases where it is much worse. Therefore the user
should be able to decide which page segmentation mode is best.
Old results for hocr, pdf or tsv now need an explicit `--psm 1` for
reproduction.
Signed-off-by: Stefan Weil <sw@weilnetz.de>
* 'master' of https://github.com/tesseract-ocr/tesseract:
Fix CID 1164579 (Explicit null dereferenced)
print help for tesstrain.sh; fixes#1469
Fix CID 1395882 (Uninitialized scalar variable)
Fix comments
Move content of ipoints.h to points.h and remove ipoints.h
remove duplicate help from combine_lang_model
Fix typo.
use tprintf instead of printf to be able disable messages by quiet option (issue #1240)
add "sudo ldconfig" to install instruction. fixes#1212
unittest: Replace NULL by nullptr
unittest: Format code
tesseract app: check if input file exists; fixes#1023
Format code (replace ( xxx ) by (xxx))
Simplify boolean expressions
Win32: use the ISO C and C++ conformant name "_putenv" instead of deprecated "putenv"
The report from Coverity Scan is a false positive.
Nevertheless the code can be rewritten and optimized
a little bit to fix that report.
Signed-off-by: Stefan Weil <sw@weilnetz.de>
The implementation for ICOORD only allows division by scale != 0.
Do the same for FCOORD by asserting that scale != 0.0f,
so undefined program behaviour will be caught.
Signed-off-by: Stefan Weil <sw@weilnetz.de>
Only print "Merging rows..." if textord_debug_blob==true (like all the other debug messages).
Otherwise, there are a lot of "Merging rows..." messages in console output.
The error message "segmentation fault" confuses most users,
so enforce a segmentation fault only in debug code.
Release code simply calls the abort function.
Signed-off-by: Stefan Weil <sw@weilnetz.de>
Wrong or old parameters in traineddata files should not terminate
the program, so make that a warning instead of a fatal error.
This fixes issue #1520.
Signed-off-by: Stefan Weil <sw@weilnetz.de>
