Classifier Fusion Based on Inner-Cluster Class Distribution

Abstract

Multiple classifier fusion is an effective way to improve the classification performance. In this paper, member classifiers are designed based on the training dataset’s different feature spaces. By utilizing ISODATA technique, various clustering results can be obtained in different feature spaces. For each member classifier (corresponding to one feature space), the given test sample is assigned to the cluster according to the distance between each cluster centroid and the test sample itself. The mass functions and the classification decision of each member classifier for the given test sample can be implemented based on the corresponding cluster’s inner-cluster class distribution. Then according to Dempster’s rule of combination, the multiple classifier fusion can be implemented. Experimental results show the rationality and efficacy of the proposed approach.