Awesome

Qaida

Qaida is a data set of 18,569 ligatures in Urdu language synthetically generated in 256 different fonts. The data set consist for 3,713,800 training images and 1,039,864 test images, each having 80x80 pixels belonging to 18,569 ligature classes. Each ligature class in the training data images rendered using 200 unique fonts while the test data contains each class rendered in 56 fonts. Fonts are kept unique across training and test set making it viable for a font-independent OCR system.

Here is an example how the data looks (each row represents a different class):

Table of contents

1. Introduction to Urdu Script
2. Get the data
3. Data usage tutorials
4. Pre-trained models

Urdu Script

Urdu is written in Arabic script in a cursive format from right to left using an extended Arabic character set. Two or more characters are joined as a single glymph to form a ligature. There are about 18,569 valid ligatures in Urdu.

Why we made Qaida?

OCR algorithms, have received a significant improvement in performance recently, mainly due to the increase in capabilities of artificial intelligence algorithms. However this advancement is not evenly distributed over all languages. Urdu is also among the languages which did not receive much attention, specially in the font independent perspective. One of the main reason for this, is intra-class variability of the letters in Urdu; unlike English and other Latin based languages, letters in Urdu are not constrained to only a single shape. Their shape changes with their position within the ligature and vary from one to four. There are almost 18,569 valid ligatures in Urdu which are to be recognized as compared to only 52 characters (excluding numbers and punctuations) in English.

Qaida is an attempt to advance the research in font independent printed Urdu test recognition. To the best of our knowledge, this is the first large scale multi-font data set for Urdu language.

Get the data

Data format

The training and test data sets are arranged in the following data structure:

train
|
├── 0               // directory name is class index
│   ├── 1.jpg
│   ├── 2.jpg
│   └── ...
|
├── 1               
│   ├── 1.jpg
│   ├── 2.jpg
|   └─── ...
|
└── ...

Mapping directory/class to ligature

Since the ligatures are in unicode format the directory names are kept as unique integers, starting from 0 to 18,568. The mapping from index to ligature can created using the mapping files present in ./data/ligatures_map for 18,569 classes and ./data/ligatures_map_2k for 2,000 classes. These mapping files can also be downloaded alongside the data set. The code for reading the mapping is as follows:

import codecs
with codecs.open('./data/ligatures_map', encoding='UTF-16LE') as ligature_file:
    ligatures_map = ligature_file.readlines()

class_idx = 18313
ligature = ligatures_map[class_idx]
print(ligature)

>>>  نستعلیق

Tutorials

• Pytorch

  for loading and training on first 2,000 classes [Warning! GPU instance required]

  for loading and training on all classes [Warning! TPU instance required]

• Tensorflow

  for loading the dataset in tensorflow 2.0 [Warning! GPU instance required]

Using the pretrained model

• Download and install anaconda for your OS
• Create the environment using ./qaida_env.yml

conda env create -f qaida_env.yml

• Activate the environment using

conda activate qaida_env

• Download the pre-trained models using the download script

cd ./src
bash src/download_models.sh
cd .. 

• Open Jupyter notebook using

jupyter notebook

• Open src/Inspect_trained_model.ipynb to see the example usage of pre-trained model

Models Stats

Contributing

Thanks for your interest in contributing! There are many ways to get involved; start with these open issues for specific tasks.

Citing Qaida

If you use Qaida dataset, trained models, or code repository in your research we would appriciate refrences to the following paper:

Large Scale Font Independent Urdu Text Recognition System. arXiv:2005.06752

Biblatex entry:

@misc{rehman2020large,
    title={Large Scale Font Independent Urdu Text Recognition System},
    author={Atique Ur Rehman and Sibt Ul Hussain},
    year={2020},
    eprint={2005.06752},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Authors

• Sibt Ul Hussain
• Atique Ur Rehman

License

Copyright (c) 2020 Atique Ur Rehman

Data licensed under CC-BY 4.0: https://creativecommons.org/licenses/by/4.0/

TODO

• End to end training tutorial