Ensemble of Randomized Linear Discriminant Analysis for face recognition with single sample per person

Linear Discriminant Analysis (LDA) has been widely used in appearance-based face recognition. However, it requires lots of training samples for each person with respect to the large dimensionality of the image space, which is difficult to collect in reality. To overcome the severe constraint of training sample deficiency, approaches based on single training sample per person (SSPP) arise in the past decades. Though making great improvements for years, these methods still suffer from low accuracy when dealing with high dimensional image features. In this paper, we develop a new variant of LDA that addresses the SSPP problem especially and apply random projections to generate extra useful training samples on an ensemble of low-dimensional subspaces. A novel extension to kernel version is also presented. We demonstrate the functionality of the proposed methods that outperform the state-of-the-arts on several benchmarks of face recognition.