Face Hallucination Using Cascaded Super-Resolution and Identity Priors.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Low-resolution facial images pose challenges for recognition and analysis.
• Existing super-resolution methods struggle with high magnification factors and preserving facial identity.

Purpose of the Study:

• To develop a novel deep learning model for hallucinating high-resolution facial images from low-resolution inputs.
• To incorporate identity priors into the super-resolution process to improve facial image quality and recognition.

Main Methods:

• A cascaded super-resolution network progressively upscales low-resolution images in 2x steps.
• An ensemble of face recognition models provides identity priors during training.
• Supervision signals are applied at multiple resolutions with identity constraints.

Main Results:

• The proposed C-SRIP model generates visually convincing high-resolution facial images from unconstrained, low-resolution inputs.
• The model demonstrates superior performance compared to state-of-the-art methods on LFW, Helen, and CelebA datasets.
• The cascaded approach with identity priors effectively enhances image quality and preserves identity information.

Conclusions:

• The C-SRIP model offers a robust solution for face hallucination, particularly at high magnification factors.
• Incorporating identity priors is crucial for generating high-fidelity facial images.
• The cascaded architecture enables effective multi-scale supervision and constraint application.