Improvements of pattern recognition by using evidence theory. Application to tag identification

The authors describe the improvements provided to a pattern recognition task by the use of evidence theory when combining different classifier results. The application of this method concerns the identification of buried metal tags detected by an eddy current sensor. These tags are characteristic of the different contents (gas, water, ...) of the buried pipes. We have developed classical, fuzzy and neural classifiers, each one giving a confidence level relative to its decision. We show that an appropriate mass distribution coupled with a classical combination rule, without any a priori knowledge, provide a more important performance improvement than that obtained by the application of a simple weighted voting method.