[Paper Review] NumtaDB - Assembled Bengali Handwritten Digits

Samiul Alam, Tahsin Reasat+2 more|arXiv (Cornell University)|Jan 1, 2018

Handwritten Text Recognition Techniques15 references16 citations

TL;DR

This paper introduces NumtaDB, a large-scale, publicly available dataset of 85,596 handwritten Bengali digits compiled from six diverse sources across Bangladesh, including students from public universities and children from educational institutes. The dataset was rigorously curated through multi-stage manual validation and standardized preprocessing to ensure high legibility and balanced representation across age, gender, and region, enabling robust benchmarking of Bengali handwritten digit recognition systems with minimal bias.

ABSTRACT

To benchmark Bengali digit recognition algorithms, a large publicly available dataset is required which is free from biases originating from geographical location, gender, and age. With this aim in mind, NumtaDB, a dataset consisting of more than 85,000 images of hand-written Bengali digits, has been assembled. This paper documents the collection and curation process of numerals along with the salient statistics of the dataset.

Motivation & Objective

• To address the lack of a large, publicly available, and unbiased dataset for Bengali handwritten digit recognition (HWDR).
• To reduce bias in training and evaluation data by ensuring diverse representation across age, gender, and geographic origin.
• To compile and curate a standardized, open-source dataset from multiple existing sources with consistent quality control.
• To ensure high legibility and accurate ground truth labels through multi-stage manual verification and outlier removal.
• To provide a benchmark dataset that supports reliable evaluation of Bengali HWDR algorithms across diverse writing styles.

Proposed method

• Collected digit images from six distinct sources: BHDDB, B101DB, OngkoDB, DUISRT, BanglaLekha-Isolated, and UIUDB, spanning multiple institutions and collection periods.
• Used marker-based image alignment with corner markers and geometric transformation to extract digits from grid-form scans, applying blob detection and shape criteria (circularity and eccentricity) to identify marker centroids.
• Applied SURF-based feature matching and square detection for markerless forms (OngkoDB), aligning scanned images to a reference template before digit extraction.
• Employed 1D summation and peak detection along X and Y axes to locate grid lines and extract individual digit bounding boxes from aligned images.
• Implemented a multi-stage manual validation process: initial grouping by digit, blind review by two evaluators, rechecking with odd/even digit separation, and final outlier skim.
• Applied strict legibility criteria: rejected digits with overwriting, incomplete strokes, grid line intrusion, or visual ambiguity, ensuring only clearly identifiable digits were retained.

Experimental results

Research questions

• RQ1 Can a large-scale, publicly available dataset of Bengali handwritten digits be assembled from diverse, existing sources with minimal demographic bias?
• RQ2 How effective is a multi-stage manual curation process in removing illegible, ambiguous, or corrupted digit samples while preserving writing style diversity?
• RQ3 To what extent does combining datasets from different institutions and age groups improve the robustness of Bengali HWDR models compared to single-source datasets?
• RQ4 Can automated digit extraction from scanned forms be reliably achieved using geometric constraints and feature-based alignment, especially when markers are absent?
• RQ5 Does the inclusion of samples from children and varied regional contributors enhance the generalization capability of HWDR systems?

Key findings

• The final NumtaDB dataset comprises 85,596 high-quality, manually verified handwritten Bengali digit images after pruning 13,552 illegible or corrupted samples.
• The dataset includes contributions from over 2,700 individuals across diverse age groups (6–28 years), genders, and geographic regions, significantly reducing demographic bias.
• The training set contains 72,044 digits (84.2% of total), while the test set includes 13,552 digits (15.8%), with a strict 85%-15% split applied across all datasets except UIUDB.
• The UIUDB dataset was fully retained in the test set due to its unique origin from mobile phone photos and paint applications, ensuring real-world variation in test data.
• The inclusion of the BanglaLekha-Isolated dataset—sourced from children and preprocessed with noise removal and aspect ratio normalization—enhanced the dataset’s representativeness of young writers.
• The final dataset maintains balanced digit distribution across all 10 classes (0–9), with training and testing sets preserving approximately equal sample counts per digit class.