OCRopus_Overview_(Published).rtf

{\rtf1\ansi\deff0\deflang2057\plain\fs24\fet1
{\fonttbl
{\f0\froman Arial;}
}
{\info
{\createim\yr2018\mo3\dy16\hr3\min58}
}

\paperw11907\paperh16840\margl1800\margr1800\margt1440\margb1440
\slmult0\ltrpar\li0
{\b\fs24
OCRopus Overview (Published)
}
\par\pard\plain
\slmult0\ltrpar\li200
{\fs24
commands
}
\par\pard\plain
\slmult0\ltrpar\li400
{\fs20
recognition
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
Tools for recognizing text on collections of pages.
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
ocropus-nlbin
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
noise and border removal
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
page rotation correction
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
grayscale normalization
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
binarization
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
usage
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
ocropus-nlbin image1.png image2.png ...
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
inputs
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
image1.png
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
outputs
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
image1.nrm.png
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
normalized grayscale image
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
image1.bin.png
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
binarized image
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
notes
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
optimized for 300dpi book pages
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
other kinds of inputs may require parameter adjustments
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
parameters
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
TBD
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
ocropus-gpageseg
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
column finding
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
text line finding
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
usage
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
ocropus-gpageseg page1.png page2.png ...
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
ocropus-gpageseg 'book/????.png'
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
inputs
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
page1.png
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
binarized or normalized page image
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
won't work well on other kinds of grayscale inputs
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
also looks for page1.nrm.png and page1.bin.png and uses them if available
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
outputs
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
page1.pseg.png
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
page segmentation file
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
page1/010001.png
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
text lines
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
page1/010001.nrm.png
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
grayscale text line (if .nrm.png page image was available)
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
notes
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
optimized for 300 dpi book pages
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
other kinds of inputs may require parameter adjustment
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
this program is really more of a placeholder and will be replaced by trainable layout analysis
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
limitations
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
no text/image segmentation; pages with complex figures will result in a lot of "noise"
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
column finding performance is variable depending on the input documents
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
text lines printed in very large fonts will get erroneously split horizontally, leading to misrecognition
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
TODO
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
improve column finding
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
add image detection / removal
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
ocropus-rpred
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
apply an neural network recognizer
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
usage
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
ocropus-rpred line1.png line2.png ...
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
inputs
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
line1.bin.png
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
will automatically expand glob patterns in its arguments
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
outputs
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
line1.txt
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
notes
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
recognizer is currently trained on UW3 and UNLV data; you can load other models with -m
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
also performs text line normalization
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
limitations
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
current models are trained on modern type fonts and flatbed scans (UW3, UNLV); performance will be worse on other kinds of inputs
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
parameters
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
-m model.pyrnn
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
a line recognizer model is just a pickled Python object
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
TBD
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
ocropus-hocr
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
outputs hOCR encoded recognition output
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
usage
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
ocropus-hocr bookdirectory -o book.html
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
notes
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
outputs hOCR format, including bounding boxes for text lines and per-line baseline information
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
also outputs font size information
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
TODO
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
lots of bug fixing and testing using the new pipeline
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
optionally output word and character bounding boxes
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
output more font-related information
}
\par\pard\plain
\slmult0\ltrpar\li400
{\fs20
training tools
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
Tools for training OCRopus to recognize new scripts and languages.
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
ocropus-rtrain
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
train a neural network recognizer
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
usage
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
ocropus-rtrain -o newmodel.pyrnn line1.png line2.png ...
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
inputs
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
line1.png line1.gt.txt etc.
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
will automatically expand glob patterns in its arguments
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
outputs
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
new recurrent neural network model
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
notes
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
transcriptions should be in Unicode
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
internally, text lines will be normalized prior to training/recognition; the normalization process may need modification for some scripts
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
parameters
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
-s 4
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
show every fourth input for debugging purposes
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
TBD
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
ocropus-gtedit html
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
line-wise correction of input text
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
usage
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
ocropus-gtedit html -o mypage.html *.png
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
inputs
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
PNG image files representing text lines
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
corresponding .txt or .gt.txt files
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
outputs
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
an HTML file containing both images and corresponding text (in text input boxes)
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
ocropus-gtedit extract
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
extract ground truth from an edited HTML file created with ocropus-gtedit
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
usage
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
ocropus-gtedit extract -o mydir mypage.html
}
\par\pard\plain
\slmult0\ltrpar\li400
{\fs20
visualization tools
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
Tools that helps with debugging and identifying the sources of recognition errors.
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
ocropus-showline
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
show text line recognition results and details
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
ocropus-showpage
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
show page recognition results and details
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
usage
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
ocropus-showpage book/0001.bin.png
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
ocropus-showpage book/0001.bin.png -o debug.png
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
TODO
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
ocropus-showpage 'book/????.bin.png' --index index.html
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
ocropus-visualize-results
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
ocropus-submit
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
submit a bug report
}
\par\pard\plain
\slmult0\ltrpar\li400
{\fs20
character-level recognizer
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
These are the tools for an older, character level recognizer.
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
recognition
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
ocropus-lattices
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
compute recognition lattices for text lines
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
usage
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
ocropus-lattices line1.png line2.png ...
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
inputs
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
line1.png
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
optionally uses line1.nrm.png
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
will automatically expand glob patterns in its arguments
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
outputs
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
line1.rseg.png
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
line1.lattice
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
optionally outputs line1.txt
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
notes
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
recognizer is currently trained on UW3 and UNLV data
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
input text should be around 300 dpi, rescale prior to recognition if necessary
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
will refuse to recognize lines that are too small or too large (override with --xheightrange)
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
examine the output/performance with "ocropus-showline image.png"; it will display all the intermediate recognition results
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
also performs text line extraction (see below)
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
limitations
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
current models are trained on modern type fonts and flatbed scans (UW3, UNLV); performance will be worse on other kinds of inputs
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
current models have not been trained on ligatures or unsegmentable pairs, leading to some digraphs being misrecognized frequently in some printing styles
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
parameters
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
-m model.cmodel
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
a character model is just a pickled Python class with a coutputs method
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
input to the character recognizer is a normalized character model (depending on the normalization mode)
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
TBD
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
ocropus-ngraphs
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
perform language modeling
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
usage
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
ocropus-ngraphs 'book/????/??????.png'
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
ocropus-ngraphs line1.lattice line2.lattice ...
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
inputs
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
line1.lattice
}
\par\pard\plain
\slmult0\ltrpar\li1600
{\fs20
a lattice produced by ocropus-lattices
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
will automatically expand glob patterns in its arguments
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
outputs
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
line1.txt
}
\par\pard\plain
\slmult0\ltrpar\li1600
{\fs20
recgonized text
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
line1.cseg.png
}
\par\pard\plain
\slmult0\ltrpar\li1600
{\fs20
aligned character segmentation for the recognized output
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
notes
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
language models are easy to train: all you need is a collection of text files, and you run ocropus-ngraphs --build over them; see below
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
parameters
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
There are many parameters that trade off the language model against characters, etc.  They can make a big difference.
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
You need to experiment (with some scripts generating various different combinations) to see what works for you.
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
Parameters will depend on how similar your unknown text is to your inputs, how good your character recognizer is, and how good its weights are
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
Language models make a distinction between weights and whether a string is in the language model at all.
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
There are a number of language models (default-2.ngraphs through default-6.ngraphs) available for download (see the web site).  There were derived from Project Gutenberg texts, as well as the UW3 and UNLV databases.
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
limitations
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
language models may not help recognition rates much, and can hurt recognition
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
language model parameters are highly dependent on both quality and content of the input
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
TODO
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
simple back-off
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
discriminative training of weights
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
character training
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
ocropus-tsplit
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
computes a tree vector quantizer for its inputs
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
ocropus-tleaves
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
computes terminal classifiers for each vector quantization bucket
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
ocropus-wtrain
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
trains a whitespace model
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
TODO
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
training of whitespace models is currently broken; this command will be replaced
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
ocropus-db
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
manipulate character databases
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
character databases are stored in HDF5 format
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
subcommands
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
cat db1 db2 ... -o db
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
concatenates one or more databases, optionally selecting classes and changing class labels
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
shuffle db -o out-db
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
take a random sample from a database or shufflle all its elements
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
tess2h5 *.tiff -o tess.h5
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
converts Tesseract box files to OCRopus HDF5 databases for training
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
predict -m model db.h5
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
evaluate a character recognizer on a database
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
cnormalize input -o db.h5
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
change a database from line-based size normalization to character based size normalization
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
zoom -f factor input -o db.h5
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
change the resolution of a database
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
ocropus-cedit
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
show and edit character databases
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
TODO
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
show and edit treevq files
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
language models
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
ocropus-ngraphs --build
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
build ngraph language models from a text corpus
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
data generation
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
ocropus-align
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
aligns text line images with ground truth
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
TODO
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
output line1.cseg-aligned.png
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
perform page level alignment (as before); maybe separate tool
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
ocropus-lattices --extract
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
extract individual characters for training
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
usage
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
ocropus-lattices --extract -o chardb.h5 image1.png image2.png
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
ocropus-lattices --extract -o chardb.h5 'book/????/??????.png'
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
ocropus-linegen
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
generate text lines for training
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
TODO
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
port over from llpy
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
ocropus-db tess2h5
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
see above
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
visualization
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
ocropus-ngraphs --print
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
usage
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
ocropus-ngraphs --print some.lattice
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
purpose
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
prints a textual representation of the recognition lattice
}
\par\pard\plain
\slmult0\ltrpar\li200
{\fs24
files
}
\par\pard\plain
\slmult0\ltrpar\li400
{\fs20
intermediate files during recognition
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
file formats
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
File formats are documented externally; please follow the links.
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
.pseg.png
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
http://goo.gl/9IfXM
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
.rseg.png
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
http://goo.gl/9IfXM
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
.cseg.png
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
http://goo.gl/9IfXM
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
.lattice
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
http://goo.gl/9IfXM
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
.aligned
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
http://goo.gl/9IfXM
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
hOCR output
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
http://goo.gl/zXNrC
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
directory layout
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
Although many OCRopus commands work on individual files, there is a standard directory layout that some tools expect.  It consists of a top level directory containing files of the form 0001.extension for each page, and containing subdirectories of the form 0001, that in turn contain files for each line.
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
book
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
0001.bin.png
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
A deskewed, binarized page image.
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
0001.nrm.png
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
A deskewed, intensity normalized page image.
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
0001.pseg.png
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
A color-coded page segmentation file.
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
0001
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
There is one of these directories for each input page.
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
Produced by the page layout analysis program (e.g., ocropus-gpageseg)
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
Each file name corresponds to the hex RGB value of the corresponding pixels in the 0001.pseg.png page segmentation file.
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
010001.bin.png
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
A binary text line image.
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
Produced by e.g., ocropus-gpageseg
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
010001.nrm.png
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
An intensity normalized grayscale text line image.
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
Produced by e.g., ocropus-gpageseg
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
010001.rseg.png
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
A raw character segmentation image.
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
Produced by ocropus-lattices
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
Some tools assume that this is also a good binarization of the input (FIXME), but really a good binarization is this image masked with the .bin.png image
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
010001.lattice
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
A recognition lattice.
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
Produced by ocropus-lattices
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
010001.cseg.png
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
An aligned character segmentation.
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
Produced by ocrpous-align and ocropus-ngraphs
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
010001.aligned
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
Aligned character data.
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
Produced by ocropus-align and ocropus-ngraphs (FIXME)
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
This encodes ligatures and pseudo ligatures using notation like "_fi_".
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
010001.gt.txt
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
Ground truth text used for training.
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
Used as input by ocropus-align
}
\par\pard\plain
\slmult0\ltrpar\li400
{\fs20
model files
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
These are the models responsible for recognition.
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
.cmodel
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
A pickled python object representing a character classifier.
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
members
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
coutputs(image) -> [(cls,prob),...]
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
Returns a list of classes and probabilities for the given input image.
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
sizemode
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
indicates how characters are expected to be scaled for this classifier
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
"perchar" - use rescaling as implemented by improc.classifier_normalize
}
\par\pard\plain
\slmult0\ltrpar\li1400
{\fs20
"perline" - use rescaling as implemented by improc.line_normalize
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
comments
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
training with "ocropus-tsplit" and "ocropus-tleaves" or other kinds of character training tools
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
primarily responsible for recognition of characters
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
cmodels may expect different size normalization of their inputs
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
mismatches between normalization and sizemode will lead to very poor recognition performance
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
the sizemode is stored in the sizemode member (see above)
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
Latin script is particularly sensitive to the size and position of characters and requires "linerel" mode; for many other scripts (Chinese, Japanese, Greek, Hebrew, etc.), the "perchar" sizemode is easier to train and use
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
If characters are trained with "linerel" mode, you need to have a .lineest model
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
.lineest
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
A pickled Python object with two methods returning polynomial models of the baseline and xline.
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
members
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
lineFit(image) -> (bl_poly,xl_poly)
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
returns two polynomials suitable for evaluation with polyval (just a list of coefficients), one for the baseline, one for the x line
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
comments
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
training with "python -m ocrolib.lineest" command
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
if the script is similar to Latin script, probably doesn't need retraining
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
Line estimators are only needed for scripts in which relative size and position of characters makes a significant difference.  For most scripts other than Latin scripts this isn't the case.
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
.ngraphs
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
A pickled Python object representing conditional probabilities.
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
members
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
all the members are private (the class is only used by ocropus-ngraphs)
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
comments
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
training with "ocropus-ngraphs" command
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
this is easy to train, and needs to be trained even for different corpora
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
recognition performance is quite sensitive to the parameters given to ocropus-ngraphs (e.g., language model weight etc.); this tricky tradeoff between language models and character recognition is true of all OCR engines
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
.wmodel
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
A pickled Python object estimating the probabilities that there is a space at a particular location.
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
members
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
setLine(image)
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
Sets the line image.
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
classifySpace(x) -> (no,yes)
}
\par\pard\plain
\slmult0\ltrpar\li1200
{\fs20
Returns the probability that there is no space, a space, for the character ending at location x in the input image.
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
comments
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
the training tools for this is currently broken; it will be replaced
}
\par\pard\plain
\slmult0\ltrpar\li200
{\fs24
library
}
\par\pard\plain
\slmult0\ltrpar\li400
{\fs20
utility
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
default
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
a few default paths
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
this is also read by setup.py
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
TODO
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
rename to defaults.py
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
toplevel
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
decorators etc. intended to be imported at toplevel (very few)
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
in particular, contains the @checks decorator, which checks arguments to Python functions
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
there is a large number of argument checks defined, relevant to image processing and OCR
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
common
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
reading and writing of images, segmentation files, etc.
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
dealing with file names and directories
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
some general simple image processing functions
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
some unicode-related utilities
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
warnings
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
checks
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
loading and instantiating components
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
TODO
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
move exceptions and annotations to toplevel.py
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
use warning, checks, etc. more consistently
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
use mkpython more consistently
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
remove old make_I... functions
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
move rect_union to sl
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
move drawing functions into separate module
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
make gt_explode/implode into Python codec
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
sl
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
utilities for dealing with Python tuples-of-slices as rectangles
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
native
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
a small library from embedding C code inside Python
}
\par\pard\plain
\slmult0\ltrpar\li400
{\fs20
image processing
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
morph
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
morphological operations to completely scipy.ndimage.morphology
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
improc
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
various image processing operations
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
lineproc
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
image processing specifically related to text lines
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
psegutils
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
image processing specifically related to page images
}
\par\pard\plain
\slmult0\ltrpar\li400
{\fs20
pattern recognition
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
patrec
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
various pattern recognition algorithms
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
mlinear
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
linear classifiers, logistic regression
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
mlp
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
MLPs and backpropagation
}
\par\pard\plain
\slmult0\ltrpar\li400
{\fs20
recognition
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
wmodel
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
white space models
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
lineseg
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
line segmentation by connected components
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
line segmentation by dynamic programming cuts
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
linerec
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
text line recognition
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
lineest
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
line geometry estimation
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
finds the baseline and xline of text lines
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
trained models of the form name.lineest
}
\par\pard\plain
\slmult0\ltrpar\li400
{\fs20
text
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
lang
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
language and script related data
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
TODO
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
merge with ligatures.py
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
ligatures
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
encoding and decoding of ligatures as integers, related functions
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
lattice
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
reads lattice files and augments them with some transitions; used for both alignment and language modeling
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
notes
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
you should usually use linerec.read_lattice and linerec.write_lattice
}
\par\pard\plain
\slmult0\ltrpar\li800
{\fs20
TODO
}
\par\pard\plain
\slmult0\ltrpar\li1000
{\fs20
change the name to something less generic
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
ngraphs
}
\par\pard\plain
\slmult0\ltrpar\li600
{\fs20
hocr
}
\par\pard\plain
\slmult0\ltrpar\li200
{\fs24
KEY
}
\par\pard\plain
\slmult0\ltrpar\li400
{\fs20
this is a command
}
\par\pard\plain
\slmult0\ltrpar\li400
{\fs20
this is an obsolete command
}
\par\pard\plain
\slmult0\ltrpar\li400
{\fs20
this is an unimplemented comand
}
\par\pard\plain
\slmult0\ltrpar\li400
{\fs20
this is a group of modules by function
}
\par\pard\plain
\slmult0\ltrpar\li400
{\fs20
this is a module in ocrolib
}
\par\pard\plain
\slmult0\ltrpar\li400
{\fs20
this is a file name in the directory hierarchy
}
\par\pard\plain
\slmult0\ltrpar\li400
{\fs20
this is a file format
}
\par\pard\plain
\slmult0\ltrpar\li400
{\fs20
click on the +/- signs to expand/collapse nodes
}
\par\pard\plain
}