Robust framework of single frame face superresolution across head pose, facial expression, and illumination variations
This paper presents a framework for single-frame face super-resolution that is robust to variations in head pose, facial expression, and illumination. The framework uses a redundant transformation with diagonal loading to model mappings between different facial factors, and local reconstruction with geometry and position constraints to incorporate image details into new factor spaces. Experimental results demonstrate that the proposed framework offers robustness when dealing with inputs with different expressions, poses, and illuminations compared to other methods, can generate higher resolution faces with better quality than tensor-based methods, and improves on previous work by producing multiple outputs with varied factors rather than a single output.
Robust framework of single frame face superresolution across head pose, facial expression, and illumination variations
- 1. ROBUST FRAMEWORK OF SINGLE-FRAME FACE SUPERRESOLUTION ACROSS HEAD POSE, FACIAL EXPRESSION, AND ILLUMINATION VARIATIONS ABSTRACT This paper presents a robust framework to solve theface hallucination problem across multiple factors, i.e., differentexpressions, head poses, and illuminations. It proposes a redundanttransformation with diagonal loading for modeling the mappingsamong different new face factors, and a local reconstructionwith geometry and position constraints for incorporating image detailsin the new factor spaces. Our proposed redundant and sparsestrategies are discussed, and the experiments indicate that it is notnecessary to adopt sparse representation in the proposed framework.The experimental results demonstrate that the proposedframework offers robustness when dealing with the inputs thathave different expressions, head poses, and illuminations comparedwith the state-of-the-art methods,can generate high-resolution faceimages with better image qualities than the hierarchical tensor basedmethod, and improves the state of the art from single oneoutput to multiple outputs with new factors.