Deep Recognition-based Character Segmentation in Handwritten Syriac Manuscripts

Ancient Syriac manuscripts represent a valuable cultural heritage for a large community, mainly in the region of the Middle East. Digitization of these manuscripts plays a major role in preserving this heritage and making it publicly available without the need to access the original paper versions. In this context, this paper presents a recognition-based algorithm for character segmentation in handwritten Syriac manuscripts. A convolutional neural network is used as a classifier for character recognition in a variable-size sliding window that spans individual word images. The classifier also provides likelihood values that are then used to determine suitable segmentation points for individual characters. Since the classifier is a core component of the segmentation algorithm, the proposed approach allows for joint character segmentation and recognition, thus providing a means for automated document indexing and content-based search and retrieval.