Simulated data for bangali text localisation in natural image

Published: 22 August 2023| Version 3 | DOI: 10.17632/h9kry9y46s.3


(1) The dataset comprises paired images, consisting of 2D simulated text on the styled image (i_s), text image (i_t), masking of text (mask_t), real background image (t_b), real sample image (t_f), text skeleton (t_sk), text styled image (t_t). (2) While generating synthetic data, we scale the text image in our experiment to a height of 64 pixels while maintaining the original aspect ratio. (3) The source dataset has background images, color, fonts, and text, whereas the synthetic dataset contains randomly 49,000 images. (4) The batch size is set to 8, and the input images are resized in size to approximately W×64 while maintaining the aspect ratio.


Steps to reproduce

INSTALLATION At first, create a Conda Environment. We choose bste_dataset as the name of the environment. conda create -n bste_dataset python==3.9.0 conda activate bste_dataset pip install -r requirements.txt Note: The GitHub repository provides concise instructions on how to reproduce all the things step-by-step.


Jahangirnagar University, United International University


Text Editing, Synthetic Image, Bengali Language, Deep Learning, Generative Adversarial Network