Combining the Classification Results of Independent Classifiers Based on the Dempster/Shafer Theory of Evidence

A procedure is defined that transform the distance measures of different classifiers into confidence values in three subsequent steps and achieves an output vector for each distance classifier which is compatible with statistically adapted classifiers.

Abstract

Recognition problems requiring flexibility are often solved with Nearest Neighbor Classifiers. This classification priniciple is very powerful even if only a few reference symbols per class are available. The main drawback is the difficulty to integrate statistical knowledge about the recognition task. This is especially obvious if different classifiers are applied to the same recognition task. Therefore a procedure is defined that transform the distance measures of different classifiers into confidence values in three subsequent steps. First each distance is transformed into an evidence function of a so called specialist, responsible only for one reference pattern. Second the theory of evidence of Dempster/Shafer is used to combine the votes of the specialists to form the evidence function of a single classifier. Third the results of the different classifiers are combined according to the same theory. The performance of the procedure is tested in the field of on-line script recognition. The classification results are compared to a former technique. The proposed technique is applicable in different fields of pattern matching and achieves an output vector for each distance classifier which is compatible with statistically adapted classifiers.