Super-resolution for human faces based on sequential images and learnt prior

In this paper, we propose to super-resolve the low-resolution (LR) facial images based on both the image sequence and the training samples. We adopt the rational of multi-surface fitting (MSF) as the foundation of the proposed method. Specifically, the confidence of the fitted surface is determined by the learned information of image derivatives. The learned information can be regarded as the probability density distributions of different derivatives at a given position. Besides, the prior estimation, which is ignored in the original MSF method, also contributes to the final results. Experiments on FERET database demonstrate the superiority of the proposed method over several state-of-the-arts.