Multi-phase attention network for face super-resolution.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• General super-resolution methods struggle with facial detail restoration.
• Prior and attribute-based methods require extra networks and data.
• Existing approaches face challenges in obtaining additional training data.

Purpose of the Study:

• To develop an effective face super-resolution method that overcomes limitations of existing techniques.
• To improve the restoration of facial structure and texture details.
• To eliminate the need for additional networks and training data.

Main Methods:

• Proposed a Multi-phase Attention Network (MPAN) for face super-resolution.
• Integrated residual attention groups (IRAG) with residual channel attention blocks (RCAB) and an integrated attention module (IAM).
• Utilized a concatenated attention module (CAM) to focus on informative layers and enhance facial texture reconstruction.

Main Results:

• MPAN demonstrated significant competitive advantages over state-of-the-art networks.
• Achieved superior performance across various scale factors and metrics (PSNR, SSIM).
• Effectively restored key facial structures like eyes and mouths.

Conclusions:

• The Multi-phase Attention mechanism significantly enhances face super-resolution.
• MPAN recovers high-resolution (HR) face images effectively without external information.
• The proposed method offers a robust solution for detailed facial image reconstruction.