Improving classification performance with focus on the complex areas

Abstract

In combining classifiers, effort is made to achieve higher accuracy in comparison with the base classifiers that form the ensemble. In this paper, we make modifications to the conventional decision template, DT, method, so that its classification performance is improved in experiments with Satimage, Image Segmentation and Soybean datasets. In our modified version, DT, an elegant strategy in classifier fusion, is used in the first stage of classification task, and in the second stage, the most misclassified classes are directed to a classifier that is specifically devoted to those classes. To identify the most misclassified classes, the confusion matrix of the output of the decision template stage is considered. Experimental results demonstrate the improved performance of the modified version by a 3% increase in the recognition rate for Satimage dataset in comparison with previously published results on Satimage dataset, a 10.57% increase in the recognition rate for Image Segmentation and 4.88% for Soybean dataset, in comparison with the conventional method. © 2010 Springer-Verlag Berlin Heidelberg.