Deciphering the Underserved: Benchmarking LLM OCR for Low-Resource Scripts

Synthetic datasets further enhanced these This study investigates the potential of Large Language Models models by mitigating the scarcity of annotated data for specific (LLMs), particularly GPT-4o, for Optical Character Recognition languages [3, 12] (OCR) in low-resource scripts such as Urdu, Albanian, and Tajik, Progress in low-resource OCR remains limited, with studies focusing with English serving as a benchmark. Using a meticulously curated on Indic scripts, while structurally complex scripts--such as dataset of 2,520 images incorporating controlled variations in text those with ligatures or modified Cyrillic alphabets--remain underrepresented length, font size, background color, and blur, the research simulates [12]. Similarly, few studies have explored OCR pipelines diverse real-world challenges. Results emphasize the limitations of for Urdu and Bengali [1, 2], highlighting the need for customized zero-shot LLM-based OCR, particularly for linguistically complex solutions. LLMs have recently emerged as a promising OCR solution, scripts, highlighting the need for annotated datasets and fine-tuned leveraging multimodal capabilities for dynamic adaptation to models. This work underscores the urgency of addressing accessibility diverse scripts and challenging visual conditions [10, 13]. Studies gaps in text digitization, paving the way for inclusive and have demonstrated LLMs' ability to process textual content from robust OCR solutions for underserved languages.