This paper presents an extension of linear discriminant analysis to higher order tensors that enables robust color object recognition. Given a labeled sample of training images, the basic idea is to consider a parallel factor model of a corresponding projection tensor. In contrast to other recent approaches, we do not compute a higher order singular value decomposition of the optimal projection. Instead, we directly derive a suitable approximation from the training data. Applying an alternating least squares procedure to repeated tensor contractions allows us to compute templates or binary classifiers alike. Moreover, we show how to incorporate a regularization method and the kernel trick in order to better cope with variations in the data. Experiments on face recognition from color images demonstrate that our approach performs very reliably, even if just a few examples are available for training.