Multi-Label Graph Cuts including Prior Knowledge for Character Segmentation
Technischer Bericht IES-2012-09. In: Jürgen Beyerer, Alexey Pak (Hrsg.), Proceedings of the 2012 Joint Workshop of Fraunhofer IOSB and Institute for Anthropomatics, Vision and Fusion Laboratory, KIT Scientific Publishing, 2012.
Recently, a character segmentation method based on graph cuts has
been introduced. In this report, we extend this method to multi-label
graph cuts. Instead of sequential segmentation of single characters
from a text line, all characters are segmented at once. The method
is especially suitable for applications where subsequent images contain
the same (known) number of characters. This prior knowledge and the
approximate positions of characters relative to the defined text
region are included in the formulation of the energy functional.
This approach provides significant segmentation quality improvement,
in particular, when characters are touching, which may happen , e.g.
due to poor print quality or scratches on the material's surface.
To illustrate the performance changes, we compare the results obtained
with this and previously known methods, based on real data.