Fusion of support vector classifiers for parallel gabor methods applied to face verification

Abstract

In this paper we present a fusion technique for Support Vector Machine (SVM) scores, obtained after a dimension reduction with Bilateral-projection- based Two-Dimensional Principal Component Analysis (B2DPCA) for Gabor features. We apply this new algorithm to face verification. Several experiments have been performed with the public domain FRAV2D face database (109 subjects). A total of 40 wavelets (5 frequencies and 8 orientations) have been used. Each set of wavelet-convolved images is considered in parallel for the B2DPCA and the SVM classification. A final fusion is performed combining the SVM scores for the 40 wavelets with a raw average. The proposed algorithm outperforms the standard dimension reduction techniques, such as Principal Component Analysis (PCA) and B2DPCA. © Springer-Verlag Berlin Heidelberg 2007.