Super-resolution for facial image using multilateral affinity function

In this paper, a patch-based super-resolution (SR) method is proposed to hallucinate facial images, where the image patches are selected and weighted based on a multilateral affinity function (MAF). Inspired by the property of human faces, we design the MAF by combining four parts, each of which is also an affinity function and inspired from different insights. The first part describes the similarity of two patches by their appearances. The second one takes the probable positions of patches into account. The third part incorporates the global information of faces by Lasso regression. The fourth one includes the information of significant facial components. Through the data consistency constraint, weights of training patches are calculated from MAF. The final SR results are obtained by the stitching of inferred HR patches and a post-processing. The experiments on two public databases demonstrate the superiority of the proposed method over some state-of-the-art methods via various criteria. The feasibility of our method in the real-world scenario is also demonstrated experimentally.