Segmentation and Normalisation in Grapheme Codebooks

The grapheme codebook is a high-performing technique for offline writer identification. This paper considers whether the de facto standards for initial grapheme extraction are optimal for both modern and historical datasets. We examine the construction and representation of the graphemes that comprise the codebook, testing three segmentation methods and two grapheme size normalisation methods on two datasets: a 93-writer IAM dataset, and a 43-writer medieval English dataset. The standard minima-split segmentation is compared to a complementary segmentation method that preserves ligature shapes, as well as the union of both these methods. Classification performance for each method is compared on a range of codebook sizes. We demonstrate that grapheme aspect-ratio is not always a writer-specific feature, and that preserving the character body shape in segmentation is more informative than preserving cursive text ligatures.