Mean polynomial kernel for face membership authentication

Face recognition techniques have gained much attention and research interests over the recent years due to their vast applications in security and authentication systems. Some of the popular approaches involve support vector machines (SVM), which can either be a binary or a multiclass classification problem, and subspace learning, where data is assumed to lie on some low dimensional manifold, such as that employing the Grassmann kernels. Recent trends involve data in the form of image sequences, hence treating them as data points in a Grassmann manifold and performing discriminant analysis in this space has been widely used. However, this technique requires determining the reduced dimensionality which has been a critical issue for such techniques. In this paper we introduce another kernel for face membership authentication with similarities to the Projection kernel, a Grassmann kernel. Using the proposed kernel, dimensionality reduction is of no concern and, thus, so is data loss. Moreover, data covariance matrices are directly exploited. Experimental results on face membership verification task show the effectiveness of the proposed kernel over the Grassmann kernels and the Grassmann Distance Mutual Subspace Method (GD-MSM).