The “Digital Maktaba LP”

Optical Character Recognition (OCR) plays a vital role in digitising and enabling access to historical records in digital libraries. Yet, OCR technologies frequently face challenges when interpreting and categorising intricate document structures, particularly in historical materials with varied layouts and languages. This initial study tackles the issue by proposing the creation of a rich, publicly available dataset of Arabic title pages, leveraging advanced Vision Language Models (VLMs) alongside OCR techniques. By extracting the first pages of each document in high resolution, our focus was on accurately identifying frontispieces and separating them from the main body of the text to enhance metadata accuracy and document discoveries within digital repositories. The Qwen-2vl-72B model was employed to determine whether each page is a frontispiece or non-frontispiece through a custom-designed prompt. Detected frontispieces will be processed with Google Vision AI to generate Ground Truth data, later reviewed by linguistic specialists before finalising the dataset. Future plans include training open source models like Kraken OCR to assess dataset utility. This novel strategy addresses current dataset gaps while boosting digital archive performance, including in projects like Digital Maktaba.