A New Class of Decision Surfaces based on the Minimization of within Class Variance

In this paper a modified class of support vector machine (SVM) inspired from the optimization of Fisher´s discriminant ratio is presented. The modified class of SVM is used in order to find decision surfaces by solving the corresponding optimization problem in arbitrary Hilbert spaces, defined by Mercer´s kernels. The effectiveness of the proposed approach is demonstrated by comparing it with the maximum margin SVM in various experiments using artificial data. Moreover, we have applied the proposed approach in the recognition of neutral expression in facial images.