Abstract
In this paper, we propose a novel approach based on multiobjective genetic algorithms for SVM classifiers applied to problems with limited training samples. It consists in injecting unlabeled samples into the training set after finding their optimal label through a genetic optimization process. Such process looks for the best chromosome which encodes the sample labels relying on two fitness functions that estimate the SVM classifier generalization performance. The chromosome is configured in such a way as to solve also the SVM model selection issue. The choice of the two fitness criteria is made so that to satisfy two requirements which are the sparseness of the SVM solution and the statistical compatibility with the available training samples. Experiments conducted on the basis of a multispectral image show the promising capability of the proposed approach to integrate unlabeled samples in the SVM classification process with significant gains of accuracy.