A flexible framework for online document segmentation by pairwise stroke distance learning

A flexible framework for handling a variety of segmentation problems in online handwritten documents is introduced. The strategy relies on single linkage clustering and a pairwise stroke distance that is globally trained for direct optimization of the segmentation. We define a variety of features that can contribute to the pairwise distance definition and show how to select a good combination of features for dealing with a new online handwritten document segmentation task. Our experiments demonstrate the validity of the method on a large range of segmentation tasks over several types of documents from various publicly available databases.