Early feature stream integration versus decision level combination in a multiple classifier system for text line recognition

This paper compares two different methods to combine feature streams to improve the performance of offline handwritten text line recognition systems. In both methods a pixel-based and a geometric feature stream are combined. The first method integrates the feature streams at an early stage whereas in the second method a combination step at the decision level is applied. In the experiments, the early integration approach outperforms the decision level combination as well as recognisers built from the individual feature streams