README Tesseract Regions

Python Libraries

tesserocr

https://pypi.python.org/pypi/tesserocr

Gives access to "several" API calls, uses C++ bindings, threading. Useful call:

def TesseractRect(self, imagedata,
                  int bytes_per_pixel, int bytes_per_line,
                  int left, int top, int width, int height):
    """Recognize a rectangle from an image and return the result as a string.

def GetRegions(self):
    """Get the result of page layout analysis as a list of
    image, box bounds {x, y, width, height} tuples in reading order.

Looks promising.

pyocr

https://github.com/openpaperwork/pyocr

Works with command via fork, or directly with Libtesseract.

pytesseract

https://github.com/madmaze/pytesseract

I can't tell what they are doing for 'boxes':

def run_tesseract(input_filename, output_filename_base, lang=None, boxes=False,
                  config=None, nice=0):
if boxes:
    command += ('batch.nochop', 'makebox')

but I'm guessing it asks to output boxes rather than define a region. It seems to output bboxes for every character it finds. Not very useful for me.

What can we specify in a config?:

if config:
    command += shlex.split(config)

pyocr

One reviewer liked this better than pytesseract.

python-tesseract

Looks unmaintained since 2014, no README

https://bitbucket.org/3togo/python-tesseract/src/9ce0abe168297513d648406be5482b52d38d883b/src/?at=master

But this says it gives access to the API, which would be useful:

https://www.quora.com/How-do-I-use-PyTesser-and-Tesseract-OCR-in-Ubuntu-with-Python

UZN files

"tesseract.exe pic1.bmp pic1.txt -psm 4" and put a pic1.uzn file in the current directory.

do pytesseract (boxes command?) or pyocr use uzn or bbox files?

tesseract.exe test.png test -psm 4" with tesseract, test.png and test.uzn in the same directory will result in a test.txt with the content This is another test

Content of test.uzn: What's the format?? 100 130 200 30 Text

I tried to use the numbers from the parse hocr for field 4:

{'4.': (485, 1107, 545, 1164),
 'Top': (571, 1107, 708, 1179),
 'Assy.': (730, 1107, 924, 1181),
 'Part': (950, 1107, 1097, 1167),
 'Number': (1121, 1109, 1406, 1169)}

tesseract GCAR1.tif --psm 4 GCAROUT with GCAR1.uzn:

1121 1109 1406 1169 Text

But got:

Number 5. Top
38-301 Li-IO
Asse
dumber 8. Sub;
10. Revision Summary: NM
ation: N/A
muel Russell

So I think my numbers for the uzn may be diff from the bbox?

With 7 (single line of text):

tesseract GCAR1.tif --psm 7 GCAR-psm7

I get:

muel Russell

Samuel Russel is way below "4. Top assy. Part Number".

https://github.com/OpenGreekAndLatin/greek-dev/wiki/uzn-format says the format is:

left top width height freetext

So I'm gussing the HOCR bbox is xmin ymin xmax ymax. So if I convert...:

1121 1109 1406-1121=339 1169-1109=60

or:

1121 1109 339 60 sometext

and run:

tesseract GCAR1.tif --psm 7 GCAR-uznfixed-psm7

we get what we exptected:

Number

Merge our bounding box to include all the field 4. Take min and max x, y of all fields in the HOCR bbox:

{'4.': (485, 1107, 545, 1164),
 'Top': (571, 1107, 708, 1179),
 'Assy.': (730, 1107, 924, 1181),
 'Part': (950, 1107, 1097, 1167),
 'Number': (1121, 1109, 1406, 1169)}

So bbox:

485 1107 1406 1181

Convert to uzn:

485 1107 921 74 field "4. Top Assy. Part Number"

and we get what we expect:

4. Top Assy. Part Number

Is there a way we can use Tesseract's own "image segmentation" to get locations of (say) outlined field boxes and such so we don't have to guess this stuff?