Page-level Annotated Text Recognition Data Assemblage

This article presents PATRA (Page-level Annotated Text Recognition Assemblage), a dataset designed to advance offline handwriting recognition for the Bengali script. The dataset comprises 996 high-resolution scanned images derived from 27 unique source paragraphs. These source texts were curated to ensure comprehensive coverage of the Bengali script, including all vowels, consonants, numerals, punctuation marks, and complex conjuncts (juktakkhor). By ensuring this script-level completeness, the dataset addresses the limitations of existing resources that often lack complex compound characters. The data is organized with an image–text mapping file in .csv format. PATRA serves as a critical resource for developing algorithms for line segmentation, document layout analysis, and holistic Optical Character Recognition (OCR), with applications in digital archiving and automated document processing.

Kaggle Dataset Link: https://www.kaggle.com/ds/6040509