Robust high performance character segmentation based on projections and graph cuts

Character segmentation is still an ongoing research topic, especially for images captured in industrial applications, where the demand for real-time capability and robustness is very high. For character segmentation, projection profiles are mostly used, which are restricted to linear cuts between characters. This method can be considerably improved by a procedure based on graph cuts to achieve nonlinear cuts as well. The experimental results on real data show that the performance is significantly improved, not only in contrast to projection profiles, but also using projections in combination with dynamic programming, and a method based on connected component analysis combined with a shortest path algorithm.