Restoration of ancient Arabic manuscripts: a deep learning approach

Abstract

This paper explores the application of modern deep learning methodologies to the restoration of highly valuable ancient Arabic manuscripts, a task of immense cultural and historical importance. Our approach meticulously guides readers through the experimental process, placing a strong emphasis on crucial components such as the selection of appropriate loss functions, the architecture of hidden layers, and the optimization techniques used. The results of our research are nothing short of extraordinary, particularly with the implementation of the proposed Modified Attention-Based Bidirectional Long Short-Term Memory (M-AB-LSTM) model, which achieved an outstanding accuracy rate of 99.50%. This work transcends traditional image enhancement techniques; it plays a pivotal role in not only making fragments of our rich cultural heritage accessible but also in ensuring the preservation of these priceless and unique artifacts for future generations. Such an effort is of profound significance to humanity as a whole. Additionally, we highlight the extensive and labor-intensive process involved in manually curating finely tuned and accurately classified datasets, which includes a comprehensive collection of 3,745 ancient Arabic manuscripts.