Skip to content

Latest commit

 

History

History
160 lines (108 loc) · 7.38 KB

README.md

File metadata and controls

160 lines (108 loc) · 7.38 KB

Angelina Braille Reader

Angelina Braille Reader is an Optical Braille Recognition system. It is designed to convert Braille text on photos into plain text.

This solution is available as

image

image

Note that these solutions use the most actual neural net model while the model for standalone installation available here is not always up to date.

General description of the solution

The solution is a web-service.

Users interact with it via a standard web browser on a smartphone or a desktop computer. Results are displayed on the screen as images and text and can be sent to the user's E-mail.

This solution can also be installed as a standalone program on a personal computer and can be used through a command-line interface.

Video presentation: https://youtu.be/_vcvxPtAzOM

This service is available at the address: http://angelina-reader.ru

Solution key features

  • Can handle images of deformed braille pages
  • Can recognize either one- or two-side Braille printouts
  • Can recognize both recto and verso sides of a page using a single image
  • Can automatically find the correct orientation of an image
  • Can process:
    • images taken on a smartphone camera directly from the application (only mobile web version)
    • image files (jpg etc.)
    • pdf files
    • zip-archives with images
  • Results can be sent to the user's e-mail
  • Can recognize Russian, English, German, Uzbek, Latvian and Greek braille texts

Limitations

  • Page image must be taken approximately from a top view
  • Light must fall from the upper side of the page. I.e. shadow of a subject placed on a page must be directed at the bottom side of the page. Top light, side light, and light from the bottom side of the page are not allowed.
  • Braille symbols must not be too small or too large. Optimally A4 page with standard braille text must occupy the whole image area.

Approaches used in the project

  • Braille symbols are detected using object detection CNN (RetinaNet https://arxiv.org/abs/1708.02002)
  • Primary network training was done using the DSBI dataset
  • Additional training data were prepared using several rounds of manual correction of results produced by CNN trained on a previous round dataset
  • At first rounds poetry texts were used, errors were found using line-by-line comparison with the original text
  • At later stages, recognition errors were found using spell-checker
  • A new annotated dataset of 360 pages of single-side handwritten and two-side printed Braille texts is prepared, including annotation of 76 paged from the dataset, provided by World AI&DATA Challenge contest. This dataset will be published later.
  • For an automatic search of correct page orientation, the page is processed in all 4 possible orientations and the orientation with the maximum presence of the most wide-spread Braille chars is selected
  • For recognizing or verso side text we use the effect, that dented points became visually convex on the inverted image. We invert an image and flip it horizontally to recognize the verso side.
  • We use a heuristic algorithm to form strings from detected symbols.
  • We translate Braille symbols into plain Russian or English text using an algorithm where Braille interpretation rules are coded.

Environment requirements

Standalone workstation requires NVIDIA GPU with at least 3GB memory (i.e. GeForce GTX 1050 3GB or better), web-server requires at least 4GB GPU memory (GeForce GTX 1050Ti or better)

OS: Ubuntu, Windows
CUDA 10.2
Python 3.6+
python packages see requirements.txt

Python path should be added to PATH.

A client requires a standard web-browser (Chrome, Firefox)

Installation

1. Install Angelina Braille Reader

git clone --recursive https://github.com/IlyaOvodov/AngelinaReader.git
cd AngelinaReader
pip install --upgrade pip
pip install -r requirements.txt

2. Download neural net model

wget -O weights/model.t7 http://ovdv.ru/files/retina_chars_eced60.clr.008

Note that these solutions uses the most actual neural net model while the model for standalone installation available here is not always up to date.

3. Install Liblouis library

Download and install Liblouis:

For Windows: update liblouis_tables_path_prefix parameter in AngelinaReader\local_config.py file with path to Liblouis tables, including trailing "/"

4. Run web app

python run_web_app.py

Windows: pip directory (i.e. <python>\Scripts) should be added to Path .
Be sure python and pip start Python3 if both Python 3 and Python 2.7 are installed.

Open http://127.0.0.1:5000 in a browser. The main page of the application should be displayed.

To access the application from Internet forward port 80 to port 5000 of the server. It is not required to test the service locally (at http://127.0.0.1:5000 address).

Usage

Using as a web service

start server: python run_web_app.py For Windows: you can use bat-file start_web_app.bat

Open page http://127.0.0.1:5000 in a browser. Follow instructions.

If some Braille symbols can not be interpreted by the application, they are displayed as ~?~.

Usage of web-application is demonstrated in a brief video: https://youtu.be/_vcvxPtAzOM and in a video presentation https://youtu.be/_vcvxPtAzOM

Command-line interface

python run_local.py [-h] [-l LANG] [-o] [-2] input [results_dir]
or, in Windows:
start.bat [-h] [-l LANG] [-o] [-2] input [results_dir]
Parameters:
input - image file (jpg, png etc.), pdf file, zip file with images or directory name.
If directory name or zip file is supplied, all image and pdf files in it will be processed.
results_dir - folder to place results in. If not supplied, the input files folder will be used. For every input file will be created files <input file>.marked.txt with results in a plain text form and <input file>.marked.jpg with plain text printed over input image.
-l <language> - input document language (default is RU). Use -l EN for English texts, -l GR for Greek etc. See languages list below. -o - switch off automatic orientation search. Sometimes auto orientation can work incorrectly (for non-typical texts or if there are many recognition errors). In such cases adjust image orientation manually and use -o option.
-2 - recognize both recto and verso sides of two-side printouts. Verso side results are stored in <input file>.rev.marked.txt и <input file>.rev.marked.jpg files.
-h - print help.

Languages: RU - Russian EN - English (grade 1) DE - German GR - Greek LV - Latvian PL - Polish UZ - Uzbek (cyrillic) UZL - Uzbek (latin)

Datasets being used

Network weights: see repository ./weights folder.

Auxiliary instruments

None.