Robust face hallucination via locality-constrained multiscale coding

Face hallucination (FH) is to produce face images with High Resolution (HR) from Low Resolution (LR) observations. Unfortunately, most existing FH methods fail to make full use of the local geometrical information, especially when the LR images are corrupted by noise. Inspired by the observation that regions with large scales can provide much useful information, in this paper we propose a Robust Locality-constrained Multiscale Coding (RLMC) based method to forecast HR face images while suppressing noise and outliers. In RLMC, a weight vector is used in the loss function to ease the effect of outliers in data representation. Furthermore, inspired by the observation that abundant local information can be exploited by jointly representing overlapping patches with multiple scales. Simultaneously encoding multiple scale patches encourages different scales to share complementary information, which admits the proposed method to generate more appropriate coefficients for super-resolution reconstruction. Experimental results verified the effectiveness of the proposed method in terms of both quantitative measurements and visual impressions. (C) 2019 Published by Elsevier Inc.