Two-stage patch-based multi-view face superresolution

In this paper, we propose a learning-based method to generate a high-resolution (HR) face in frontal view from a low-resolution (LR) face in an arbitrary pose. This HR virtual face (HRVF) method is based on two stages of pixel-structure learning. In the first stage of our algorithm, initially estimated HR frontal-view images are generated from non-frontal-view LR input images, based on a patch-based learning method. In the second stage, the estimated frontal-view image will be used to search for similar faces from the interpolated LR frontal-view face database. The targeted HR frontal-view face image is then constructed based on the local patches of the HR faces of the corresponding LR face images in the database. Experiments show that the proposed algorithm can produce a better performance than existing methods.