Low resolution face recognition with pose variations using deep belief networks

In practice face recognition sometimes encountered by low resolution (LR) face images with varying poses, which degrade the performance significantly. To address this problem, we propose an approach that applies deep belief network (DBN) to handle the non-linearity caused by pose variations. The manifold assumption states that point-pairs from high resolution (HR) manifold share the topology with the corresponding LR manifold. Inspired by this assumption, we learn the relationship between HR manifold and LR manifold by sending both HR images and LR images to a deep architecture. High performance is achieved in the experiment on ORL and UMIST, in which great facial pose variations present.