Multilayer selection-fusion model for pattern classification

Individual classification models are recently challenged by combined pattern recognition systems. In such systems the optimal set of classifiers is first selected and then combined by a specific fusion method. Large and rough search space formed from performances of various combinations of classifiers makes the selection process very difficult and often leads to selection overfitting, degrading generalisation ability of the system. In this work a novel design of multiple classifier system is proposed, which recurrently uses multiple selection and fusion processes applied at many layers to a population of best combinations of classifiers rather than the individual best. On the particular implementation with evolutionary search algorithms and majority voting, the improvement of the system's generalisation performance is demonstrated experimentally and explained theoretically.