Optical character recognition in Linux

Introduction

This is not just a review of existing OCRs (we’ll only talk about three ) and not an installation guide (although the installation will be described). This article was created in order to understand what and how can really recognize the Russian and English languages ​​in Linux.

A few words in order to understand the essence of the described processes.
OCR - optical character recognition.
Technology is needed to digitize printed documents; some use OCR for automation purposes (for example, to recognize captcha or to protect against spam bots).

OCR in Linux

I repeat once again: there will be considered programs that recognize the Russian language. Under Linux, there are several OCRs designed for working with the Latin alphabet, there are specialized complexes that work only with Hebrew, for example, all this does not apply to our topic.
In fact, only three products will be discussed: Cuneiform , Tesseract and Finereader Engine . All of them by themselves provide only a console interface, although for the first two GUI is sufficiently developed.

I use Debian Squeeze, but I will often provide links to source codes and explain package builds (you can use the repositories on notesalexp.org or repositories of your distribution kit - I just give an example of the build).
The topic will be revealed in the following order:
')
1. Installing OCR for Linux (3 engines), installing them.
2. OCR CLI comparison with examples.
3. GUI for OCR, their comparison.
4. A small test online-OCR.
5. Conclusions and some predictions and suggestions.

Install OCR for Linux

Cuneiform

A page about a project on wikipedia .

Declared features: support for multiple languages, preserve the formatting of the original document, output to txt, hocr, html, recognize faxes and texts printed on a matrix printer.

We will look at two real ways to use Cuneiform under Linux: native and with the help of Wine (there is a need for this, below you can see for yourself).

1. Native Cuneiform

Launchpad section
Sources

Let's start the installation.
Download the sources and unpack them.
Then everything is standard (see the readme.txt bundled with the source). Go to the source directory and execute sequentially:

mkdir builddir cd builddir cmake -DCMAKE_BUILD_TYPE=debug .. make sudo checkinstall sudo ldconfig 

Is done.

2. Installation under Wine.

The advantage of this method is that we immediately get the original functional GUI. The Wine version is not important (Cuneiform worked under Wine 1.0). The only feature: it is necessary to specify a new substitution for the msvcrt library in the Wine settings.
Distribution is available at this link .

Tesseract.

A page about a project on wikipedia .
Page on Google Code .

Declared features: support of multiple languages, output in txt and hocr, training the program with my own examples (I will not consider this, work as is), use of the configuration file for a specific sample.

I briefly and loosely describe this Readme .

Allow dependencies before installation:

 sudo apt-get install autoconf automake libtool libpng12-dev libjpeg62-dev libtiff4-dev zlib1g-dev 

Tesseract also depends on libleptonica-dev version not lower than 1.67. In Squeeze, this package is outdated, so I had to collect it.
We get the source code , unpack and build:

 ./autobuild ./configure make sudo checkinstall sudo ldconfig 

Now we get the Tesseract source , unpack it and go to the directory with them.
Next, perform:

 ./autogen.sh ./configure make sudo checkinstall sudo ldconfig 

Tesseract is installed. We will receive packages for language recognition: Russian and English - and unpack them into the tessdata directory ( / usr / local / share / tessdata by default).
You can work.

FineReader Engine

An announcement can be viewed here .

How to get? We go here , carefully read and ask for a trial (limit of 100 recognitions). You can ask in Russian .

Installation is simple: download , run under the root abbyyocr.run and follow the text instructions.

Declared features: support for multiple languages, various encodings, work with passwords, page numbering, recognition of tables, bar codes, texts printed on a matrix printer, typewriter, gothic fonts, etc., output in txt, rtf, html, xml, xls.

Rubyquet +

Since I started talking about CLI OCR, I’ll mention CLI Rubyquet + for Tesseract and Cuneiform.
I did not test it (you can do this) - the same CLI for the above-mentioned OCR is available by itself.

OCR CLI comparison

* CLI - command line interface - command line interface (“console”).

I warn you immediately: this section is quite voluminous. If you want to do without unnecessary details and a lot of letters - I recommend not to disclose spoilers.

Remember: I do not claim absolute objectivity of comparison. You may have other results and other conclusions.

I will designate the test criteria.

Of course, the perfect result should be one hundred percent recognition of all characters, formatting and pictures. However, in practice, text recognition is the most popular. The necessary formatting and addition of text with images is able to be made by the user during post-processing.
To assess the quality of recognition, I will introduce the following criteria (although I will depart from them):

The criterion “ Wrong words ” - incorrectly recognized words (from one wrong character in a word up to the complete absence of a word) for simplicity of calculation is the most important criterion.

The criterion of “ Incorrect characters ” - incorrectly recognized characters when it is impossible to use the first criterion (extra characters, punctuation, etc.).

The criterion “ Formatting errors ” is used to determine the quality of work with tables, figures, definition of writing with bold and italic.

For each sample and program, the percentage of correctly recognized words will be calculated, which will be a fundamental factor. The number of words in the sample I will take on the result of recognition in Finereader.

Now about the samples.
What do we most often recognize? Either the documentation passed through the scanner, or a photograph of the document. Naturally, with different quality and resolution (for OCR, scanning is recommended as at least 300 dpi, therefore we compare scanned samples to 200dpi, 300 dpi and 600dpi; for photos, we use photography with 2MP and 5MP quality). In addition, some samples will be tables and pictures.

I will give recognizable images in the form of links to them (their direct presence in the article will only interfere). The recognition result for the first link will be available on Google Docs, the second link with the mark “Original” - in its original form on Dropbox

In order not to clutter up the review, instead of the paths of the original image and the resulting text, I will write INPUT and OUTPUT, respectively.

Sample number 1 (numbered list).

Sample number 2 (scan English textbook).

Sample number 3 (photographed English textbook).

Sample No. 4 (recognition of tables and pictures of the scanned image).

Sample No. 5 (scanned textbook “Metal constructions”).

Sample # 6 (scanned O'Henry story page).

Sample №7 (scanned page of the book "The Moment of Truth").

Sample №8 (Pangram with different fonts).

GUI for Linux.

* GUI - graphical user interface - a graphical interface (“windows and buttons”).

Yagf

Official page.

Dependencies when building: libaspell-dev and libqt4-dev version 4.5 or later. Qt 4.5 and aspell are needed to run (see the documentation that comes with the source).

Installation (start in the source directory):

 mkdir builddir cd build dir cmake ../ make sudo checkinstall 

Yagf is well localized, can get an image from the clipboard, from a file, from a scanner, import pdf, allows you to align the image.

In the Yagf settings, you can switch between Tesseract and Cuneiform. Yagf can produce batch recognition (all imported images) or recognition of a specific text area.
The only significant disadvantage I consider is the impossibility of setting additional parameters for scan engines, i.e. similar to the command line (OCRFeeder, discussed below, does not have this drawback).

Cuneiform-qt

The project began and ended its active life in April 2009 as part of the Altlinux project. Cuneiform-Qt provides a simple GUI for Cuneiform.

Since I didn’t expect anything special from this GUI, I decided to limit myself to installing the ready-made package version 0.1.1-1 (the latest source version 0.1.2 - the development was not far).

The GUI, by the way, turned out to be very interesting - when it was saved in RTF, the recognized text turned into a sequence of several Latin characters repeated hundreds of times, arranged in a column one letter wide. In a plain text file, the save is normal.

We conclude: this GUI is useless.

KBookOCR

This was announced on Habré “killer Finereader'a”, which is a superstructure over Cuneiform.
Official blog of the author.
Deb package.

Unfortunately, in dependencies pulls a piece of KDE.Version 2.2, planned over a year ago, implies Tesseract support, but there seems to be no progress.
For the review, I used Kubuntu 12.04 in Virtualbox.

Version 2.1 of this software can get an image from a file or by scanning, display the result in html or open it in a text editor. To distinguish blocks for recognition, KBookOcr, unlike Yagf, does not know how.

Conclusion: KBookOcr loses Yagf in terms of functionality and is intended only for KDE.

OCRFeeder

GUI for Cuneiform, Tesseract and a pair of other OCRs that do not support Russian. In deb-packages are presented only very old versions, so we will collect from the source code .

Immediately there is an irritation: Readme is written to version 0.3, the current version is 0.7.1. The file structure has changed, there is no required setup.py. But there is ./confugure
Dependencies:

 sudo apt-get install python-pygoocanvas ocrad unpaper python-gtkspell python-enchant sane python-imaging-sane 

Further, it turns out that:
Your intltool is too old. You need intltool 0.35.0 or later.
The old intltool-debian package is installed by default : we put the intltool version 0.41 package from the repository .

We collect:

 ./confugure make sudo checkinstall 

The set of functions is standard: open / scan, save / open in the editor. There are several bonuses: unpaper is supported, the settings of the recognition engine allow you to transfer to it in text form any supported parameters (for example, the choice of the recognition language is done this way); You can select text blocks.

Ocropus

http://code.google.com/p/ocropus/ - either GUI or Tesseract CLI.
Following the instructions, I tried to build the latest version, but python said about some source line:
SyntaxError: invalid syntax
I don’t want to go into the code, I conclude that the product is more dead than alive.

gImageReader

Tesseract GUI.
Take the deb package from here .

gImageReader is as simple as possible: it allows you to open an image, adjust brightness and contrast on the go and recognize the text area or all at once.

Tesseract-gui

Project page.
Packages lie here .
I refused to recognize anything.

Conclusion about the GUI :
The only real advantage presented by the GUI may be the recognition function of individual text blocks (some cannot and this).
Therefore, I believe that the existing Linux OCR GUIs are not functional. If we choose among them, then in fact the only acceptable will be only two: Yagf and OCRFeeder. At the same time, the Yagf community is used and mentioned much more often.
Try and leave your impressions.

Online-OCR test

A small list of online-OCR .
Online OCR is very much, and, in principle, they are all implemented on the three engines listed above: Cuneiform , Tesseract and Finereader .
Since the OCR data falls under the topic of this article, we consider a couple of them, at the same time we will draw interesting conclusions.

1. Finereader Online .
finereader.abbyyonline.com/ru
Clearly based on Finereader Engine 9 (or maybe 10?) I did not try, I am sure of high quality.
Allows you to recognize 10 pages per day for free.

2. New OCR .
www.newocr.com
Free, does not require registration and has no restrictions.
A very interesting resource that will allow you to draw conclusions about the ultimate applicability of Cuneiform and Tesseract . Load the images with which these two systems have problems, and look at the result.

Sample No. 3, 5MP without flash

Recognition with Cuneiform :
0006.cun.newocr.txt
( Original )

Recognition with Tesseract :
0006.tes.newocr.txt
( Original )
The online recognition results are obviously better.

Sample # 7

Recognition with Cuneiform :
0014.cun.newocr.txt
( Original )

There are not enough hyphens and dashes, however, there are only 2 errors in the words. Excellent result.

Recognition with Tesseract :
0014.tes.newocr.txt
( Original )
6 errors in words, a few extra characters - but Tesseract in my image could not recognize normal text. Perhaps the most interesting test result.

Here's a great example of using free OCR: the site creators explicitly applied image pre-processing and (possibly) text post-processing (adjustments using dictionaries or something like that). And in this form, free OCRs can already compete with Finereader.

Conclusion

We considered three OCR systems that can work with Russian and English.

Without a doubt, FineReader Engine v9.0 showed the best result. It perfectly recognizes both scanned and photographed images. However, its minimum cost is € 149 for a license for 12,000 recognitions per year - do you need this?

Free OCR: Cuneiform and Tesseract - by themselves can only adequately process scanned images with uniformly distributed across the field brightness and high contrast.
Both engines coped well with the English-speaking sample, while problems arose with the Russian text - in general, during the test, it turned out that the results of the work of free OCR "float" from sample to sample.
At the same time, CuneiformV12, running under Wine, coped with Russian texts better than the native version under Linux.

Interesting was the fact that, on the examined samples, the optimal image quality was 200 dpi - with a higher number of dots per inch, the quality of recognition began to deteriorate.

The Cuneiform and Tesseract GUIs are many, but they don’t bring real value.

On the example of FineReader and the online system New OCR, it is clearly seen that a normal functional OCR must necessarily exist in conjunction with an image preprocessing system and a text post-processing system based on vocabulary control of the results.
At the same time, this technology for free engines has actually been implemented (in New OCR, recognition is very high quality).
It is enough to share such technology or implement it yourself - and the free OCR will be on the same level with ABBYY products - after that you can talk about writing a good GUI.

To summarize: free OCRs do not provide a stable result even on scanned images, Finereader costs money - it is easier for a simple user to use online services.
Well, if you are going to use OCR on a different scale, there’s a different conversation: you have to pay or spend time adjusting the results manually.

PS If you have ready-made ideas or scripts for preparing an image for recognition - share in the comments. Everyone will be interesting and useful.