Ensemble of Log-Euclidean Kernel Svm Based on Covariance Descriptors of Multiscale Gabor Features For Face Recognition

Face recognition (fr) is a challenging computer vision task due to various adverse conditions. local features play an important role in increasing the recognition rate of an fr method. in this direction, the covariance descriptors of gabor wavelet features have been one of the most prominent methods for accurate fr. most existing methods rely on covariance descriptors of gabor magnitude features extracted from single-scale face images. this study proposes a new method named multiscale gabor covariance-based ensemble log-euclidean svm (mgcov-elsvm) for fr that uses the covariance descriptors of gabor magnitude and phase features derived from multiscale face representations. mgcov-elsvm begins by producing multiscale face representations. gabor magnitude and phase features are derived from the multiscale face images in the second stage. after that, the gabor magnitude and phase features are used to generate covariance descriptors. finally, covariance descriptors are classified via a log-euclidean svm classifier, and a majority voting technique determines the final recognition results. the experimental results from two face databases, orl and yale, indicate that the mgcov-elsvm outperforms some recent fr methods.