Learning Spatial Attention for Face Super-Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• General super-resolution methods struggle with detailed face recovery in low-resolution images.
• Existing deep learning face super-resolution models often require additional labeled data for multi-task learning and are limited to lower resolutions.
• Current methods face limitations in generating high-resolution (e.g., 512x512) face images.

Purpose of the Study:

• To develop an effective and efficient face super-resolution method capable of recovering detailed facial structures.
• To introduce a novel network architecture that adaptively focuses on key facial features.
• To achieve high-resolution face image generation without relying on extensive manually labeled data.

Main Methods:

• Proposed a novel SPatial Attention Residual Network (SPARNet) incorporating Face Attention Units (FAUs).
• Introduced a spatial attention mechanism within residual blocks to adaptively focus on salient facial structures.
• Extended SPARNet to SPARNetHD using multi-scale discriminators for high-resolution (512x512) output generation.

Main Results:

• SPARNet effectively captures key facial structures even in very low-resolution images (16x16), as shown by attention map visualizations.
• Quantitative evaluations demonstrated superior performance over state-of-the-art methods across metrics like PSNR, SSIM, identity similarity, and landmark detection.
• SPARNetHD produced high-quality, high-resolution outputs and showed good generalization to real-world low-quality face images.

Conclusions:

• The proposed spatial attention mechanism enhances the efficiency and effectiveness of face super-resolution.
• SPARNet and its high-resolution variant SPARNetHD represent a significant advancement in generating detailed and high-resolution face images.
• The method demonstrates strong potential for applications requiring high-fidelity face reconstruction from degraded inputs.