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

Name	Content	Classes	Examples	Size	Link	MD5 Checksum
train.tar.xz	training set images	18,569	3,713,800	2.6 GBytes	Download	90ffe6411c5147ecc89764909cc6395a
test.tar.xz	test set images	18,569	1,039,864	461 MBytes	Download	847a146ecd9fc2db6e62a38eea475db6
ligature_map	index to ligature mapping	18,569	18,569	195.4 KBytes	Download	0c1b2e60b1c751d1a14c5eb90fec745e
train_2k.tar.xz	training set images	2,000	400,000	279 MBytes	Download	847a146ecd9fc2db6e62a38eea475db6
test_2k.tar.xz	test set images	2,000	112,000	79 MBytes	Download	847a146ecd9fc2db6e62a38eea475db6
ligature_map_2k	index to ligature mapping	2000	2,000	21.3 KBytes	Download	37bbd4e44ae486dbb5d7e98801811ae4
train_200.tar.xz	training set images	200	40,000	24.1 MBytes	Download	a42b6a78a2f73d826b7b8ccbdaf5a60b
test_200.tar.xz	test set images	200	11,200	5.9 MBytes	Download	bc0aa5b0307d5a6e122acc2767d25c04
ligature_map_200	index to ligature mapping	200	200	1.4 KBytes	Download	d8c38d3398b97549204d5d2c9a8b13ce

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

Name	Backbone	Seen (200) fonts Accuracy	Unseen (56) fonts Accuracy	Size	Link	MD5 Checksum
QRN18	ResNet18	84.2	75.5	165 MBytes	Download	b64e527360d21f55d35f0c684d6cad06

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