SMAjram: A Large-Scale Synthetic OCR Dataset for Punjabi Shahmukhi (Perso-Arabic) Script

This dataset is designed for the development, training, and evaluation of Optical Character Recognition (OCR) systems for the Shahmukhi (Perso-Arabic) script, with emphasis on character-level recognition, dataset scalability, and comparative validation of machine learning and deep learning models.

This dataset (Shahmukhi Database - Ajram) comprises an augmented collection of Punjabi Shahmukhi keyboard character images, including context-specific alphabets, numerals, symbols, and other characters. The images, all in (.png) format with a size of 100 x 100 pixels, are organized into one primary folder and several secondary folders. The primary folder contains unique (.png) images featuring six different font sizes and 50 fonts designed explicitly for Perso-Arabic text. The dataset contains images with various augmentations, including text rotations, blurring, changes in background and contrast, and outline effects. Each secondary folder includes one copy of each image from the primary folder for comparison, while all other images in the dataset are unique.

It’s important to note that scripts such as Urdu, Arabic, and Persian are subsets of the Punjabi Shahmukhi script in terms of their keyboard characters. Although scripts such as Sindhi, Balochi, Pushto, Saraiki, and Kashmiri have unique characteristics, they can also use this dataset to develop algorithms for their scripts. The secondary folders enable researchers to evaluate their algorithms on this dataset for different parameters, including alphabets, characters, numerals, font sizes, and typefaces, across various difficulty levels tailored to their specific requirements.

The dataset includes a ground-truth file set detailing the features of each image in the file names. Additionally, the secondary folders offer enhanced versions of these images. There are 240 Shahmukhi keyboard characters, six font sizes, and 50 fonts, resulting in 72,000 PNG images in the primary folder (main folder). In primary folder 2, five rotation angles are applied to each of the 72,000 images, resulting in a total of 360,000 images. Secondary folders include blurred images (1440k), text-background gray-level variations (4680k), images, and text-border-outline-background variations, totaling 1,33,20k images. There are 1,98,72,000 (19.872 million) images in this dataset, all in PNG format (100 x 100 x 3).

There are two sets of the same dataset, one with Shahmukhi labels and the other with Latin labels (written in Roman).