Automatic Recognition of Ancient Arabic Manuscript Authors Using Deep Learning

Abstract

This study investigates the effectiveness of deep learning techniques in the automatic identification of authors of ancient Arabic manuscripts. The motivation lies in preserv- ing Arab cultural heritage and supporting academic research in the field of palaeog- raphy. Given the absence of ready-made, task-specific datasets, a custom dataset was manually compiled, comprising 10 authors, each represented by 20 manuscript images. This dataset served as a starting point to assess the feasibility of authorship classifica- tion based on handwriting patterns. To prepare the data for model training, a series of preprocessing steps and data augmentation techniques were applied to improve input quality and reduce overfitting risk. The study implemented a simple Convolutional Neural Network (CNN) model and evaluated three pre-trained architectures: VGG16, InceptionV3, and ResNet50. The results demonstrated that the custom CNN model achieved a test accuracy of 90%, while VGG16 and InceptionV3 reached 95% due to the advantages of transfer learning. In contrast, ResNet50 performed less effectively, achieving 75% accuracy, likely due to its complexity and higher data requirements. The study confirms that deep learning models can capture distinctive handwriting features even from relatively small datasets. These findings support the potential of AI-based systems in historical manuscript analysis and suggest future directions, such as expanding the dataset, integrating metadata, and developing practical tools for automated authorship attribution in Arabic script manuscripts.