Optical character recognition with Tesseract

From XPUB & Lens-Based wiki

dependencies

materials

2 hi-res (300dpi) PDFs

  • a usual PDF: in English, with a common font
  • a unusual PDF: in a non-latin script, or with an unusual font or with abundant images

install

Tesseract (with languages you will be using)

  • Mac brew install tesseract --all-languages
  • Debian/Ubuntu: sudo aptitude install tesseract-ocr
    • See what language packages are available with: sudo aptitude search tesseract-ocr-
    • install language packages: sudo aptitude install tesseract-ocr-nld tesseract-ocr-port tesseract-ocr-spa here I am installing Dutch, Portuguese, Spanish

poppler-utils whic include tools such as pdftotext and pdftohtml

  • Mac brew install poppler
  • Debian/Ubuntu: sudo aptitude install poppler-utils

imagemagick

  • Mac brew install imagemagick
  • Debian/Ubuntu: sudo aptitude install imagemagick

pdftk

  • Mac brew install pdftk
  • Debian/Ubuntu: sudo aptitude install pdftk

PDF

  • with text layer
  • without text layer

To find out the difference you can try to select the PDF's text in a PDF viewer. Only if the text layer is present will you be able to select it.

If it contains a text layer you can use pdftotext command-line application (from poppler-utils) to convert the PDF to text


Tesseract

Tesseract was originally developed at Hewlett-Packard Laboratories Bristol and at Hewlett-Packard Co, Greeley Colorado between 1985 and 1994, with some more changes made in 1996 to port to Windows, and some C++izing in 1998. In 2005 Tesseract was open sourced by HP. Since 2006 it is developed by Google.

https://github.com/tesseract-ocr/tesseract/blob/master/README.md



The blog post that announced Tesseract's availability seems designed to nudge more people into action—current shortcomings are described in some detail, including trouble handling grayscale and color sources, a lack of page layout analysis features, and the program's only recognizing English-language documents. It reads less like an admission of faults than as a list of feature requests.

Bylund, Anders. 2006. ‘Google Releases Open-Source OCR Tool with HP Special Sauce’. Ars Technica. 5 September 2006. https://arstechnica.com/information-technology/2006/09/7664/.


one page prototype

Getting 1 page from PDF file with PDFTK burst 

pdftk yourfile.pdf burst

Or use imagemagick 

convert -density 300 Typewriter\ Art\ -\ Riddell\ Alan.pdf Typewriter-%03d.tiff

Chose page you want to convert

Convert PDF to bit-map using imagemagick, with some options to optimize OCR 

convert -density 300 page.pdf -depth 8 -strip -background white -alpha off ouput.tiff
  • -density 300 resolution 300DPI. Lower resolutions will create errors :)
  • -depth 8number of bits for color. 8bit depth == grey-scale
  • -strip -background white -alpha off removes alpha channel (opacity), and makes the background white
  • output.tiffin previous versions Tesseract only accepted images as tiffs, but from v3 Tesseract accepts JPEG, PNG, TIFF, BMP, PNM GIF, and WEBPformats are accepted


OCR 

tesseract output.tiff -l eng output

Will generate the file output.txt

-l is the option for language (English is the default)

Improving image quality

There are several image transformation that will improved the OCR results


See Tessearct page on improving quality of images for OCR

Suggestion: implement of this options into Imagemagick convert command.


Advanced

LibGuide Image.jpg

language

Lists all tesseract languages available in your system. 

tesseract --list-langs

If OCRing a document with more than one language Tesseract can use also more than one 

tesseract output.tiff -l eng+spa output

multipages

Tiff files can be multi-page images. Hence if we use the prevoious IM command to convert a PDF to a TIFF, if the PDF is multi page, so will be it TIFF. Which Tesseract should handle. 

$ tesseract TypewriterArt.tiff TypewriterArt
Tesseract Open Source OCR Engine v3.03 with Leptonica
Page 1 of 8
Page 2 of 8
Page 3 of 8

Another option is providing Tesseract with a text file containing the path/filename to each image in sequence:

list.txt: 

p001.tiff
p002.tiff
p003.png

tesseract list.txt output

segmentation

Page Segmentation Mode (-psm) directs the layout analysis that Tesseract performs on the page.

By default, Tesseract automates the page segmentation, but does not perform orientation and script detection.

From Tesseract man page: 

       -psm N
           Set Tesseract to only run a subset of layout analysis and assume a certain form of image. The options for N are:

               0 = Orientation and script detection (OSD) only.
               1 = Automatic page segmentation with OSD.
               2 = Automatic page segmentation, but no OSD, or OCR.
               3 = Fully automatic page segmentation, but no OSD. (Default)
               4 = Assume a single column of text of variable sizes.
               5 = Assume a single uniform block of vertically aligned text.
               6 = Assume a single uniform block of text.
               7 = Treat the image as a single text line.
               8 = Treat the image as a single word.
               9 = Treat the image as a single word in a circle.
               10 = Treat the image as a single character.

searchable PDF

tesseract input.tiff output -l eng pdf



hocr

Tesseract 3.0x supports a hocr option, which creates hocr file.

HOCR is an HTML+XML (XHTML) file consisting of recognized words and their coordinates. 

tesseract input.tiff -l eng output hocr


The HOCR file contains all pages as ocr_page elements. with attribute that contains the following fields :

  • ppageno: The physical page number
  • image: The relative path (from the HOCR file) to the page image
  • bbox: The dimensions of the image

class='ocr_page

The OCRed text is atomized into text elements of different magnitude, such as:

  • paragraph "ocr_par"
  • line "ocr_line"
  • word "ocrx_word"


HOCR tools

using hocrjs

We will use User Script instruction with Tampermonkey.

Installing:

  • open Firefox
  • go to FF addons and search for Tampermonkey
  • install it
  • Browse to unpkg.com/hocrjs/dist/hocr.user.js
    • click "Install". It will install the script in your browser Tampermonkey
    • click the Tampermonkey and go to the "Dashboard". hocr-viewer should be enabled

Create an horc file with tesseract

Note: in this process will be more convenient to use a png or jpg input file, as the browser will not display a tiff.

Run tesseract to produce a hocr (language and segmentation options can also be used ) 

tesseract inputfilename.png inputfilename hocr

You have got a inputfilename.hocr

View the hocr int the Firefox

  • change its extension from .hocr to .html
  • open the .html file in firefox

hocr-viewer will automatically load

hocrjs-2.png

Editing and correcting

hocrjs does not support editing :(

A solution would be to use FF inspector to change the content of the HOCR, but the HTML inspector changes are not saved, even if we are working with a local html file :((

So the only option is to do the editing in a plain text editor :(((


converting the HOCR

To PDF:

It makes sense to use the position information and and plain-text content to create a text-based PDF.

hocr-pdf an application from hocr-tools is a possibility, but I only managed to create corrupted and empty PDFs

The following thread suggests using pdfbeads or HocrConverter.

HocrConverter showed the best results, but failed when including (-I) the page image in the PDF 

python HocrConverter/HocrConverter.py -I -i pg_0012.hocr -o pg_0012.pdf pg_0012.png

python HocrConverter/HocrConverter.py -h 

HocrConverter

Convert Files from hOCR to pdf

Usage:
  HocrConverter.py [-tIcbmnrV] [-q | -v | -vv] [-i <inputHocrFile>] [-f <inputTtfFile>] (-o <outputPdfFile>) [<inputImageFile>]...
  HocrConverter.py (-h | --help)

Options:
  -h --help             Show this screen.
  -t                    Make ocr-text visible
  -i <inputHocrFile>    hOCR input file
  -o <outputPdfFile>    pdf output
  -f <inputTtfFile>     use custom TTF font
  -I                    include images
  -c                    use full line text
  -b                    draw bounding boxes around ocr-text
  -n                    don't read images supplied in hocr-file
  -m                    do multiple pages in hocr and output pdf
  -r                    take hOCR-image sizes as reference for size of page
  -V                    vertical Inversion ( for ocropus: false, for tesseract: true )
  -q | -v | -vv         quiet ( only warnings and errors ) | verbose | very verbose = debug

Creating new fonts: Training

New fonts can be added to Tesseract through a training process.

Fonts are

The process of training for v.3 is complicated, but here are links for a few resources that can guide you in the process

Tesseract needs to know about different shapes of the same character by having different fonts separated explicitly.


tessdata/ dir, where data files can be found, can be found on Debian at /usr/share/tesseract-ocr/tessdata

If the dir happens to be located elsewhere you can use the following commands to find it: 

cd /
sudo find -type d -name "tessdata"

box output

The box file output consists of a plain-textfile containing x,y coordinates of each letter it found along with what letter it thinks it is

In cases where the input is a standard text, with a standard font, the result are not bad.

But when dealing with unusual fonts or hand-written scripts Tesseract has the possibility to train it.

Tesseract needs a 'box' file to go with each training image. The box file is a text file that lists the characters in the training image, in order, one per line, with the coordinates of the bounding box around the image.

https://github.com/tesseract-ocr/tesseract/wiki/Training-Tesseract-%E2%80%93-Make-Box-Files

Characterset.png 

convert -density 300 wafer.pdf -depth 8 -strip -background white -alpha off wafer.tiff

tesseract wafer.tiff wafer makebox

Edit the box file with [moshpytt https://code.google.com/archive/p/moshpytt/] 

./moshpytt.py

Boxeditor.png

Boxmaker is a JavaScript online box editor

Artistic research

Reverse OCR by http://reverseocr.tumblr.com/

tumblr_nekvu9mCc01tif66co1_250.png

Kindle Scanner by Peter Purgathofer

We are human beings! by Silvio Lorusso

human-thumb.jpg

related software

tesseract-ocr front-ends gImageReader: Debian install 

sudo aptitude install gimagereader

HOCR

Retrieved from "https://pzwiki.wdka.nl/mw-mediadesign/index.php?title=Optical_character_recognition_with_Tesseract&oldid=147056"