One important step for character recognition is character segmentation, since this mainly influences the quality of the characters and thus the classification result. There are several difficulties in character segmentation, e.g.if characters in a line of text are too close together then the segmentation algorithm may merge these characters, on the other hand characters may fall apart if a character is not connected with respect to the segmentation direction. To deal with such kinds of wrong segmentation, a graph based approach is introduced in this technical report. First experiments show that the approach is promising, but still has some shortcomings.