Rethinking Prior-Guided Face Super-Resolution: A New Paradigm With Facial Component Prior

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Facial priors like parsing maps and landmarks aid face super-resolution (FSR) by providing structural information.
• Existing FSR methods struggle with inaccurate priors from low-quality images and only use priors during reconstruction, not generation.

Purpose of the Study:

• To propose a novel pre-prior guided approach for face super-resolution.
• To improve face reconstruction performance by generating higher quality low-resolution (LR) face images rich in high-frequency (HF) information.

Main Methods:

• Extracting facial prior information directly from high-resolution (HR) face images.
• A component hybrid method fusing HR facial components with LR backgrounds to create new LR images (LRmix) using HR facial parsing maps.
• Developing a component hybrid network (CHNet) to learn the mapping from LR to LRmix for practical application.

Main Results:

• The proposed method successfully generates HF information-rich LR face images.
• The CHNet effectively learns the LR to LRmix mapping for real-world applicability.
• Experimental results demonstrate significant improvements in face reconstruction performance for FSR.

Conclusions:

• The novel pre-prior guided approach overcomes limitations of existing FSR methods.
• Generating enhanced LR images with fused HR components improves the overall FSR pipeline.
• This technique offers a promising direction for advancing face reconstruction quality.