Estimating Human Legibility in Historic Manuscript Images - A Baseline

Simon Brenner, Lukas Schügerl, Robert Sablatnig

Abstract: For accessing degraded historic text sources, humanist research increasingly relies on image processing for the digital restoration of written artifacts. A problem of these restoration approaches is the lack of a generally applicable objective method to assess the results. In this work we motivate the need for a quality metric for historic manuscript images, that explicitly targets human legibility. Reviewing previous attempts to evaluate the quality of manuscript images or the success of text restoration methods, we can not find a satisfying solution: either the approaches have a limited applicability, or they are insufficiently validated with respect to human perception. In order to establish a baseline for further research in this area, we test several candidates for human legibility estimators, while proposing an evaluation framework based on a recently published dataset of expert-rated historic manuscript images.