Semantic Face Hallucination: Super-Resolving Very Low-Resolution Face Images with Supplementary Attributes.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Existing face hallucination methods struggle with ambiguity, leading to distorted details and incorrect attributes.
• Low-resolution (LR) inputs lack high-frequency details present in high-resolution (HR) counterparts.

Purpose of the Study:

• To develop a novel face hallucination method that reduces ambiguity and improves the accuracy of super-resolved facial attributes.
• To enhance the super-resolution of tiny, unaligned face images with a large upscaling factor.

Main Methods:

• Developed an attribute-embedded upsampling network comprising an autoencoder with skip-connections and deconvolutional layers.
• Incorporated facial attribute vectors into residual features at the autoencoder's bottleneck.
• Utilized a discriminative network to enforce attribute accuracy during training.

Main Results:

• Successfully super-resolved tiny (16x16 pixels) unaligned face images with an 8x upscaling factor.
• Significantly reduced ambiguity and uncertainty in the one-to-many mapping problem.
• Achieved superior face hallucination results compared to state-of-the-art methods.

Conclusions:

• Supplementing residual features with attribute information effectively reduces ambiguity in face super-resolution.
• The proposed attribute-embedded network offers a promising approach for high-fidelity face hallucination.