Segmentation of brush strokes by saliency preserving dual graph contraction

Brush strokes are segmented from works of art by a combination of filtering and grouping. Filtering yields local evidence for crossings and lines. Grouping is done on two levels of scale and abstraction. The first level is a dual pair (Gu, Gu) of attributed plane graphs, the vertex and edge attributes of which are derived from the filtering. The result of the grouping on this level is given by a topological minor Gtop of Gu. The derivation of Gtop from Gu is done by dual graph contraction, i.e. by parallel steps, each of which involves only local operations on Gu and Gu. This step is shown to preserve connections via most salient paths. On the second level consecutive edges of Gtop are grouped to strokes which are consistent with our model of strokes from superimposed brush moves. Experimental results are presented for portrait miniatures.