We develop a common theoretical framework for combining classifiers which use distinct pattern representations and show that many existing schemes can be considered as special cases of compound classification where all the pattern representations are used jointly to make a decision. An experimental comparison of various classifier combination schemes demonstrates that the combination rule developed under the most restrictive assumptions—the sum rule—outperforms other classifier combinations schemes. A sensitivity analysis of the various schemes to estimation errors is carried out to show that this finding can be justified theoretically.

This article presents an off-line method for recognizing handwritten digits. Structural information and quantitative features are extracted from images of isolated numerals to be classified by a hybrid multi-stage recognition system. Feature extraction starts with the raw pixel-image and derives more structured representations like line-drawings and attributed structural graphs. Classification is done in two steps: a) the structural graph is matched to prototypes, b) for each prototype there is a neural classifier which has been trained to distinguish digits represented by the same graph-structure. The performance of the described system is evaluated on two large databases (provided by SIEMENS AG and NIST) and is compared to other systems. Finally, the combination of the described system and a TDNN classifier is discussed. The experimental results indicate that there is an advantage in using structural information to enhance an unstructured neural classifier. 1.