Stylistic Similarities in Greek Papyri Based on Letter Shapes: A Deep Learning Approach

This paper addresses the issue of clustering historical handwritings according to similarity in absence of metadata on date or style. While releasing a new dataset called AlphEpMu, it proposes to use SimSiam deep neural network model to evaluate similarity between images of individual characters from three Greek letter categories (alpha, epsilon and mu). Two specimens of a given letter category are defined as similar if they come from the same piece of manuscript (penned by a single writer). Similarity is then computed between pairs of images of the same letter category from different manuscripts. Last, scores of the three letter categories are merged to express similarity between the manuscripts. Applied to Greek Literary papyri, this approach is proved useful to paleographers since it allows organizing a complex group of 72 manuscripts (AlphEpMu-72) into a meaningful network but also spotting micro phenomena of similarity which can help explaining the general evolution of handwritings.