A multiresolution approach to on-line handwriting segmentation and feature extraction

We present an approach to on-line handwriting segmentation into elementary strokes. The segmentation is achieved by exploiting curvature information extracted from the electronic ink at different level of resolution. Such information is then combined into a saliency map, through which the segmentation points are eventually found. Information from the saliency map is also used to select the optimal resolution to be used for describing the curvature of each stroke. Experiments conducted by encoding stroke curvature information into strings have shown that similar strings are associated with similar parts of the original ink, independently of the actual word they belong to.