Ol-Chiki Digit Recognition (OLDIR) Dataset

All contributors had received formal primary or secondary education in the Santhali language, ensuring that the collected handwriting reflected authentic Ol-Chiki writing styles. Data acquisition was conducted through two complementary modes: pen-and-paper (PP) and a stylus digital writing pad (SDWP). Ninety-three volunteers contributed samples using an A4-sized pen-and-paper form containing a 10×10 grid, in which each Ol-Chiki digit was written ten times.

The remaining three volunteers provided samples using a VEIKK S640 digital writing pad. The digital samples included black, red, green, and blue handwritten digits. One left-handed participant was deliberately included to capture additional handwriting variability.

For digitization, we scanned the paper-based forms at 300 DPI using an HP DeskJet 2135 scanner. We extracted individual digit images and systematically labelled them by volunteer number, digit class, repetition number, and acquisition mode. Each digit was mapped to a separate class from 0 to 9.

A structured preprocessing pipeline was applied to improve sample consistency and image quality. This pipeline included Otsu binarization, non-local means denoising, image reshaping, and skew correction using the enhanced Piecewise Covering by Parallelogram (ePCP) technique. A two-level verification procedure was then used to identify unsuitable, blurred, poorly contrasted, malformed, or visually ambiguous samples.

Samples with severe extra strokes, broken arcs, loops, or unrecognizable handwriting were excluded, and minor defects in digital samples were manually corrected when appropriate. Following preprocessing and quality verification, the final OLDIR dataset was constructed as a class-balanced collection of 10,000 images, comprising 1,000 samples for each of the 10 Ol-Chiki digit classes. Approximately 93% of the samples originate from the pen-and-paper mode and 7% from the digital writing-pad mode.

This combination introduces variations in stroke thickness, skewness, writing style, acquisition modality, and contributor demographics, making OLDIR suitable for benchmarking conventional deep-learning, transfer-learning, and vision-transformer-based Ol-Chiki digit-recognition systems.