[Paper Review] BanglaWriting: A multi-purpose offline Bangla handwriting dataset

M. F. Mridha, Abu Quwsar Ohi+3 more|arXiv (Cornell University)|Nov 15, 2020

Handwritten Text Recognition Techniques10 references4 citations

TL;DR

This paper introduces BanglaWriting, a high-quality, manually annotated offline Bangla handwriting dataset comprising 260 individuals' single-page writings with word-level bounding boxes and Unicode labels. It contains 21,234 words, 32,787 characters, 5,470 unique vocabulary items, and includes realistic variations like overwriting and strikes, enabling advanced applications in OCR, writer identification, and handwriting generation.

ABSTRACT

This article presents a Bangla handwriting dataset named BanglaWriting that contains single-page handwritings of 260 individuals of different personalities and ages. Each page includes bounding-boxes that bounds each word, along with the unicode representation of the writing. This dataset contains 21,234 words and 32,787 characters in total. Moreover, this dataset includes 5,470 unique words of Bangla vocabulary. Apart from the usual words, the dataset comprises 261 comprehensible overwriting and 450 handwritten strikes and mistakes. All of the bounding-boxes and word labels are manually-generated. The dataset can be used for complex optical character/word recognition, writer identification, handwritten word segmentation, and word generation. Furthermore, this dataset is suitable for extracting age-based and gender-based variation of handwriting.

Motivation & Objective

• To create a comprehensive, manually annotated offline Bangla handwriting dataset that supports diverse machine learning applications.
• To include realistic handwriting variations such as overwriting, strikes, and mistakes for robust model training.
• To enable research in writer identification, optical character recognition, and handwriting segmentation using real-world writing samples.
• To explore age- and gender-based handwriting variations through a demographically diverse dataset.
• To provide raw and pre-processed image data with standardized JSON annotations for compatibility with deep learning frameworks.

Proposed method

• Handwritten pages were collected from 260 individuals across eight districts in Bangladesh using A4 paper and standard pens.
• Images were captured via scanners (HP Scanjet 2400) and smartphone cameras (Xiaomi Redmi 6/7) to include real-world imaging conditions.
• Each image was manually cropped and annotated using Labelme software to generate word-level bounding boxes and Unicode text labels.
• Annotations were stored in JSON format with 'label', 'points', 'imagePath', 'imageHeight', and 'imageWidth' fields for machine learning integration.
• Supplementary Python and OpenCV scripts were developed to reduce lighting variations and background noise in raw images.
• The dataset was split into 'raw.zip' (unprocessed images) and 'converted.zip' (pre-processed images) to support diverse experimental needs.

Experimental results

Research questions

• RQ1 How can a manually annotated, multi-purpose Bangla handwriting dataset improve the performance of offline OCR and writer identification systems?
• RQ2 To what extent do realistic handwriting artifacts like overwriting and strikes affect the robustness of handwriting recognition models?
• RQ3 Can demographic factors such as age and gender be reliably inferred from handwriting features in a Bangla context?
• RQ4 How do imaging conditions (scanned vs. smartphone-captured) impact the quality and usability of handwriting datasets?
• RQ5 What is the feasibility of using this dataset to train generative handwriting models for Bangla script?

Key findings

• The dataset contains 260 unique handwriting samples, with 21,234 words and 32,787 characters, including 5,470 unique Bangla vocabulary items.
• The dataset includes 261 instances of overwriting and 450 handwritten strikes or mistakes, enhancing realism for model robustness testing.
• All bounding boxes and labels were manually generated, ensuring high annotation quality compared to automated labeling in similar datasets.
• The dataset was collected from writers of diverse ages (8–80 years) and genders, enabling demographic-based handwriting analysis.
• The dataset includes 52 scanned images and 208 smartphone-captured images, reflecting real-world imaging variability.
• Supplementary pre-processing scripts were successfully developed to reduce lighting and noise effects in smartphone-captured images, improving data usability for deep learning.