גהאד אלצאנע

Paleography studies the writing styles of manuscripts and recognizes different styles and modes of scripts. We explore the applicability of hard and soft-labeling for training deep-learning models to classify Hebrew scripts. In contrast to the hard-labeling scheme, where each document image has one label representing its class, the soft-labeling approach labels an image by a label vector. Each element of the vector is the similarity of the document image to a certain regional writing style or graphical mode. In addition, we introduce a dataset of medieval Hebrew manuscripts that provides complete coverage of major Hebrew writing styles and modes. A Hebrew paleography expert manually annotated the ground truth for soft-labeling. We compare the applicability of soft and hard-labeling approaches on the presented dataset, analyze, and discuss the findings.