Finding text in natural scenes by figure-ground segmentation