Using methods, computer-readable storage media, and apparatuses for computer-implemented processing, an image of handwritten text may be segmented into a disjoint component image corresponding to individual glyphs connected by ligatures. The disjoint component image is skeletonized into a grid graph, and a connected path traversing the ligatures is determined. The disjoint component image is segmented into non-overlapping segments based on connected graphs corresponding to edges in the disjoint component image, where sets of adjacent non-overlapping segments correspond to the individual glyphs. Glyph geometry may be varied by obtaining an ensemble of glyph representations, each characterized by measurable geometric glyph properties. For each geometric glyph property, target values are obtained from ensemble-wide statistical distribution functions, and the target values are used to transform a base glyph representation into a target glyph representation.