MSFSR: A Multi-Stage Face Super-Resolution With Accurate Facial Representation via Enhanced Facial Boundaries

Yunchen Zhang, Yi Wu, Liang Chen; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020, pp. 504-505

Abstract


The majority of Face Super-Resolution (FSR) approaches apply specific facial priors as guidance in super-resolving the given low-resolution (LR) into high-resolution (HR) images. To improve the FSR performance, various kinds of facial representations were explored in the past decades. Nevertheless, there remains a challenge in estimating high-quality facial representations for LR images. To address this problem, we propose novel facial representation - enhanced facial boundaries. By semantically connecting the facial landmark points, enhanced facial boundaries retain rich semantic information and are robust to different spatial resolution scales. Based on the enhanced facial boundaries, we design a novel Multi-Stage FSR (MSFSR) approach, which applies the multi-stage strategy to recover high-quality face images progressively. The enhanced facial boundaries and the coarse-to-fine supervision facilitate the facial boundaries estimation process in producing high quality facial representation. The one-time projection of the FSR task is decomposed into multiple simpler sub-processes. In these ways, the MSFSR estimates a more robust facial representation and achieves better performance. Experimental results indicate the superiority of our approach to the state-of-the-art approaches in both qualitative and quantitative measurements.

Related Material


[pdf] [supp]
[bibtex]
@InProceedings{Zhang_2020_CVPR_Workshops,
author = {Zhang, Yunchen and Wu, Yi and Chen, Liang},
title = {MSFSR: A Multi-Stage Face Super-Resolution With Accurate Facial Representation via Enhanced Facial Boundaries},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2020}
}