Pro-UIGAN: Progressive Face Hallucination From Occluded Thumbnails.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Reconstructing high-resolution (HR) faces from occluded, low-resolution (LR) thumbnails presents significant challenges.
• Existing methods often struggle with artifacts and blurriness when dealing with occluded facial data.

Purpose of the Study:

• To develop an effective method for hallucinating authentic HR faces from occluded thumbnails.
• To improve the quality and usability of facial data for downstream computer vision tasks.

Main Methods:

• Propose a multi-stage Progressive Upsampling and Inpainting Generative Adversarial Network (Pro-UIGAN).
• Exploit facial geometry priors for replenishing and upsampling occluded faces (8x upsampling from 16x16 pixels).
• Utilize a novel cross-modal attention module for joint feature learning between facial appearance and landmark features.

Main Results:

• Pro-UIGAN iteratively estimates facial geometry priors and acquires non-occluded HR faces.
• The coarse-to-fine approach significantly reduces artifacts and blurriness in hallucinated faces.
• Achieved visually pleasing completed HR faces from occluded LR inputs.

Conclusions:

• Pro-UIGAN demonstrates superior performance in hallucinating HR faces from occluded thumbnails.
• The method enhances facial data quality, facilitating downstream tasks like face alignment, parsing, recognition, and expression classification.