A novel method for example-based face super-resolution

Abstract

Face super-resolution is the specific super-resolution problem based on the property of facial images, which reconstructs a high-resolution facial image from a low-resolution input. Based on the observation that faces are made up of several relatively independent parts such as eyes, noses and mouths, we propose an example-based face hallucination framework which includes correlation-constrained non-negative matrix factorization (CCNMF) algorithm and High-dimensional Coupled NMF (HCNMF) algorithm. Compared with existing approaches, the proposed CCNMF algorithm can generate global face more similar to the ground truth face by learning a parts-based and localized representation of facial images. Moreover, residue compensation by using HCNMF can learn the relation between high-resolution residue and low-resolution residue to better preserve lost high frequency details. Experimental results verify the effectiveness of our method.