Abstract: Based on the original Covariance Discriminative Learning method, a bidirectional dimension reduction was applied on those covariance matrices, which lied on a Riemannian manifold. Combining these covariance matrices with an efficient Riemannian metric, i.e., log euclidean distance (LED), a kernel function that mapped the covariance matrix from Riemannian manifold to Euclidean space for classification was derived. As a result of dimension reduction of covariance matrices, this method improved accuracy of classification and reduced the complexity of computation. The results were observed through experiments on standard datasets.

Key words: discriminant analysis, riemannian metric, face recognition, covariance matrix, object classification