Journal of Pattern Recognition Research

The question of how we can exploit the ability to combine different learning entities is fundamental to the core of automated pattern analysis and dictates contemporary research efforts in the field of decision fusion. While the broad class of information fusion methods is constantly enriched, their proper consideration on the basis of data or information distribution lacks a common framework and develops around ad-hoc methods that cannot justify the overall effectiveness of fusion methods. In this context, the present work aims at uncovering analogies between decision fusion methods and established primary classifiers. Such correspondence of specific fusion methods to base classifiers allows us to utilize knowledge from the field of data mining as to summarize and model the statistical performance of combiners and possibly provide best practices and optimality criteria for their use. As case studies, we focus on two main categories of classifiers, namely distance and discriminant-function based, when applied to the problem of classifier fusion. The Decision-Templates fusion method is examined as a representative distance based technique and compared with the Support-Vector-Machine scheme as representative of discriminant-function hyper classifiers. Based on statistical performance measures, we advocate the use of SVMs for decision fusion as an efficient and extensible framework that can be adapted to specific application domains.