Learning to Super-Resolve Face Images via Dual-Domain Multi-Scale Feature Interaction

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Deep learning significantly advanced Face Super-Resolution (FSR).
• Existing CNN and Transformer models struggle with facial structure integrity and multi-scale texture capture.
• Limitations stem from architectural constraints and rigid receptive fields in current FSR methods.

Purpose of the Study:

• To propose a novel dual-domain feature interaction method for Face Super-Resolution.
• To address limitations in facial structure preservation and multi-scale texture details.
• To introduce the Spatial-frequency Multi-scale feature Learning Network (SMLNet).

Main Methods:

• Developed a dual-branch architecture for FSR.
• Frequency branch focuses on global structures and high-frequency details.
• Spatial branch preserves fine-grained local texture patterns.
• Introduced a Multi-scale Spatial-frequency feature Interaction Module (MSIM) for feature aggregation.
• MSIM integrates Multi-scale Feature Extraction Block (MFEB) and Spatial-Frequency feature Interaction Module (SFIM).

Main Results:

• SMLNet demonstrated superior performance in quantitative experiments across multiple datasets.
• Qualitative analyses confirmed the effectiveness of SMLNet in FSR.
• Evaluations on real-world images validated the method's practical applicability.
• The dual-domain approach successfully captured both global structures and local textures.

Conclusions:

• SMLNet significantly outperforms state-of-the-art FSR methods.
• The proposed dual-domain feature interaction effectively enhances facial image quality.
• SMLNet offers a promising solution for improving low-resolution facial images.