Predicting the scribe behind a page of medieval handwriting

This paper addresses the issue of attributing pieces of medieval handwriting to scribes known from other examples of writing. The system is applied to manuscript page images and performs extraction and comparison of letter shapes. Letters and sequences of connected letters are identified by means of connected component labeling. This is followed by further splitting into letter-size pieces. The prediction process makes use of a dataset with instances of four letter types (b, d, p, and q), taken from manuscript pages with known scribes. Nearest neighbor classification is used for letter-level prediction of scribe (and grapheme). The image features capture the distribution of foreground, as it appears after a binarization step. Cosine similarity is used as the similarity metric. The system predicts the scribe behind a page by means of a voting procedure taking the highest-scoring letter-level hits for a page as its input. Evaluated on codicological units from five different scribes the system reached an accuracy above 99% for four of them and 87% for the fifth one. p, 557 q, 557 q, 557 p, 557 p, 557 p, 557 p, 557 d, 557 p, 557 d, 557 b, 557 p, 557 q, 557 d, 112 d, 557 d, 557 b, 557 b, 557 q, 186 b, 557 Table 1: An example of the top 20 letter hits for a manuscript page (Cod. Sang. 557, p. 2). Scribe prediction gives two errors.