Learning Structural Similarities from Handwriting on Papyri - An Application to Scribe Characterization

Computational analysis of historical manuscripts has remained an interesting as well as a challenging research problem for the pattern recognition community. The recent advancements in different areas of machine learning and the successful completion of joint projects between computer scientists and paleographers have resulted in increased research interest in ‘digital paleography’. Among various tasks the rich collections of historical documents offer, this study focuses on characterizing of writer from handwriting on papyri. We formulate the problem in a similarity learning framework and exploit Siamese neural networks to compare an input pair of writing samples. The highly degraded images of handwriting on papyri are binarized using a deep learning-based binarization technique. Patches of handwriting are subsequently extracted by sampling the complete image and pairs of samples from the same and different writers are produced. These positive and negative pairs are employed to fine-tune a Siamese neural network with different popular architectures as the backbone model. An experimental study on the samples in the GRK-Papyri dataset with 97 writing samples reports a test accuracy of 62% on DenseNet-121 and 69% with omniglot-trained CNN. The preliminary findings are quite promising and can be employed in writer identification, verification, and retrieval frameworks.