Position-Patch Based Face Hallucination via High-Resolution Reconstructed-Weights Representation

Position-patch based face hallucination methods aim to reconstruct the high-resolution (HR) patch of each low-resolution (LR) input patch independently by the optimal linear combination of the training patches at the same position. Most of current approaches directly use the reconstruction weights learned from LR training set to generate HR face images, without considering the structure difference between LR and the HR feature space. However, it is reasonable to assume that utilizing HR images for weights learning would benefit the reconstruction process, because HR feature space generally contains much more information. Therefore, in this paper, we propose a novel representation scheme, called High-resolution Reconstructed-weights Representation (HRR), that allows us to improve an intermediate HR image into a more accurate one. Here the HR reconstruction weights can be effectively obtained by solving a least square problem. Our evaluations on publicly available face databases demonstrate favorable performance compared to the previous position-patch based methods.