Large scale style based dating of medieval manuscripts

摘要: In this paper we propose a novel approach for manuscript dating based on shape statistics. Our goal was to develop strategy well suited large scale effort where heterogeneous collections of thousands manuscripts could be automatically processed. The proposed method takes the gray image as input, then uses stroke width transform and statistical model gradient find ink boundaries. Finally, distribution over common shapes, quantified using context descriptors, is produced each image. binarization-free, rotational invariant requires minimal segmentation. We evaluate our work 10000+ collection "Svenskt diplomatariums huvudkartotek", consisting charters from medieval period todays Sweden. images, originally intended web viewing, were low quality had compression artifacts. Due unsupervised feature learning regression, dated with median absolute error below 19 years even though only used 5% labels in estimator training.