A robust classification method using combined classifiers in a nonstationary environment

In this paper, we present a robust data classification method based on an ensemble of feature subspaces. The objective is to improve or preserve the performances of a decisional system in the case of perturbations due to noise or sensor degradation. The proposed method is to combine a set of classifiers each of which is established in the corresponding feature subspace resulting from projections of the initial full-dimensional space, expecting that most of them are not impaired. The counterpart of the expected robustness is a performance decrease for non-impaired data. In this context, three classification methods are tested, One-class SVM, Kernel PCA and Kernel ECA, to study the robustness of the final decision. The results obtained in textured image segmentation demonstrate that our approach is efficient in a nonstationary environment.