Constructing effective SVM ensembles for image classification

This paper proposes a novel method for constructing SVM ensembles for achieving improved image classification performance. The use of an SVM ensemble, instead of a single classifier, normally aims at reducing the training complexity and improving the performance. SVM ensembles are constructed by combining a set of base classifiers. While the traditional combination methods primarily consider an individual classifier's performance on the training data, the proposed method also considers its generalization ability. With our proposed method, upon training a set of base classifiers, we estimate an optimal weight for each classifier via solving a quadratic programming problem. Then the weights can be used to combine the base classifiers to form an SVM ensemble. To reduce the classification complexity, we propose an intelligent method for filtering out the weak classifiers to obtain a small subset of the relatively strong classifiers for a simplified SVM ensemble. Experiment results show that our proposed approach outperforms other published methods. When compared to the traditional SVM ensemble method which combines all the base classifiers, the proposed method can develop simplified SVM ensembles which may achieve even better accuracy.