A Hierarchical Dataset and Deep Learning Evaluation for Ancient Arabic Manuscript Segmentation

Abstract

The complexity associated with the recognition of ancient Arabic manuscripts lies in the various styles used in calligraphy, the use of ligatures, and the extent of degradation in the documents. This investigation focuses on the implementation of hierarchical segmentation at the line, word, pseudo-word, and character levels. For this purpose, a domain-specific dataset was developed using a variety of sources and was manually annotated to incorporate the various styles and configurations used in the documents. Normalization, denoising, and enhancement techniques were used for preprocessing the dataset. Three architectures of deep learning models, YOLOv8, Faster R-CNN, and Mask R-CNN, were used to test the dataset developed in this case. From the results obtained from this experiment, it was evident that YOLOv8 was effective at the line level, while Mask R-CNN was effective at the character level. However, Faster R-CNN was competitive at the word level, but cursive scripts presented a problem for this model. This investigation’s contribution lies in the development of a hierarchical dataset and a comprehensive benchmarking test, which together contribute to the development of sustainable OCR technology and the preservation of Arabic manuscript heritage. To provide a more comprehensive evaluation, we additionally benchmark two segmentation-specific baselines (U-Net and DeepLabV3+) at the character level, and report YOLOv8 metrics as mean ± standard deviation over three random seeds. A striking finding is that architecture choice can matter more than dataset size at fine granularity levels: Faster R-CNN achieves 70.2 AP50 on pseudo-words where YOLOv8 reaches only 30.9 ± 23.1, due to input resolution constraints. We position this work as a dataset-and-benchmark contribution: the primary deliverable is a publicly released, hierarchically annotated corpus (CC BY 4.0), accompanied by a reproducible evaluation protocol