Fixed and trained combiners for fusion of imbalanced pattern classifiers

In the past decade, several rules for fusion of pattern classifiers' outputs have been proposed. Although imbalanced classifiers, that is, classifiers exhibiting very different accuracy, are used in many practical applications (e.g., multimodal biometrics for personal identity verification), the conditions of classifiers' imbalance under which a given rule can significantly outperform another one are not completely clear. In this paper, we experimentally compare various fixed and trained rules for fusion of imbalanced classifiers. The experiments are guided by the results of a previous theoretical analysis of the authors. Linear, order statistics-based, and trained combiners are compared by experiments on remote-sensing image data and on the X2M2VTS multimodal biometrics data base.