2016-11-14

Tesseract OCR

Tesseract is an open source Optical Character Recognition (OCR) Engine, available under the Apache 2.0 license. It can be used directly, or (for programmers) using an API to extract typed, handwritten or printed text from images. It supports a wide variety of languages.
— Tesseract Wiki at Github

Docker

I’m keeping a docker repository with a Debian system with tesseract and some other handy utilities installed. Unfortunately is not available at DockerHub so if you want to use it to follow this entry just follow these steps:

Clone repository.

git clone https://github.com/mariogarcia/docker.git

Go to the tesseract directory

cd docker/tesseract

Build the docker image

./bin/build.sh

Run image

docker run -it mgg/tesseract

That will open a tmux session where you can execute all the commands I’m using for this entry.

Basics

To practice a little bit I’m using spanish president Mariano Rajoy public records found at http://www.congreso.es

All records related to congressmen official incomes, properties and taxes are available from http://www.congreso.es in pdf documents. Unfortunately the way these documents are fulfilled are different from one another.

Tesseract doesn’t read PDFs

In order to enable Tesseract to read PDF documents you have to convert them to images. The easiest way to do so is by using imagemagick. You should make sure to convert pdf to image with the best quality you could:

convert -density 290 image.pdf image.png

Where XX is >72 such as 288 (which is 4x). If the resulting image is too big, then you can do:

convert -density 288 image.pdf -resize 25% image.png

Where resize=25% or larger when density=288

ls -l
total 288
-rw-r--r-- 1 dev dev 196746 Nov 13 21:49 document.pdf
-rw-r--r-- 1 dev dev  17657 Nov 13 22:01 output-0.png
-rw-r--r-- 1 dev dev  17003 Nov 13 22:01 output-1.png
-rw-r--r-- 1 dev dev  16590 Nov 13 22:01 output-2.png
-rw-r--r-- 1 dev dev   9819 Nov 13 22:01 output-3.png
-rw-r--r-- 1 dev dev  10569 Nov 13 22:01 output-4.png
There is a very wellknown script to improve converted text image called TEXTCLEANER

First attempt

I’m getting the first output to get document headers:

tesseract output-0.png header

That would create a header.txt output with all the recognized data. Unless you’re processing a book, most of the form-like documents could end up as a non-sense result.

cat header.txt

51





C.DTP 319 07/07/2016 10





CORTES GENERALES XII LEGISLATURA
DECLARACIDN1 DE BIENES Y RENTAS DE DIPUTADOS Y SENADORES2

Nombre y apellidos
MARIANO RAJOY BREY

Estado civil Régimen econémico matrimonial

CASADO GANANCIALES
...

Still you can use this processing for indexing purposes. However if you happen to need to process a set of documents with certain structure you can make use of a uzn file.

UZN

uzn is a simple text file format for describing sections of a scanned image. The migneuzn tool outputs in this format for its segmentation.
— https://github.com/OpenGreekAndLatin/greek-dev/wiki/uzn-format

For this example I’m using a UZN file to get only the first output headers in a more guided way. I will tell tesseract where to get every field I’m interested in.

The uzn file is like a csv file with the following fields separated by tab/spaces:x, y, width, height, tag.

output-0.uzn
140 302 600 66 name
143 416 454 66 civil state
776 614 824 52 city

In order to make tesseract to be aware of the uzn file you have to make the uzn file to have the same name of the processed file and also use a certain segmentation mode (-psm 4):

tesseract output-0.png result_with_uzn -psm 4

And the content of result_with_uzn.txt is:

MARIANO RAJOY BREY

CASADO

MADRID ""n"—Tv…- r—- … ñ" “ “

Not bad, but I still need a little bit of tuning to get Tesseract to recognize these fields without the current noise, but hey this is just a preliminary research.

My overall impression is that Tesseract is a great project but requires from you a deeper knowledge before getting acceptable results. Also much of the processing has to do with image processing before even using Tesseract: contrast, bluring, quality…​ I will problably continuing looking into it to know more about this interesting tool.

References