S2A: Scale-Attention-Aware Networks for Video Super-Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) are prevalent in video super-resolution (VSR).
• Existing VSR methods often overlook intermediate feature correlations, hindering model expressiveness.
• Limited feature representation impacts the quality of super-resolved videos.

Purpose of the Study:

• To propose a novel Scale-and-Attention-Aware Networks (SAA) for enhanced video super-resolution.
• To address the limitations of feature correlations in intermediate features for VSR models.
• To improve the feature expression capabilities of VSR networks.

Main Methods:

• Developed the Scale-and-Attention-Aware Networks (SAA) architecture.
• Applied differential attention mechanisms to varying temporal-length streams.
• Introduced a Criss-Cross Channel Attention Module (C3AM) for spatial and channel attention integration.
• Explored attention mechanisms on separate feature streams.

Main Results:

• Demonstrated the superiority of the SAA network on public VSR datasets.
• Achieved significant improvements in both quantitative and qualitative metrics compared to state-of-the-art methods.
• Validated the effectiveness of integrated spatial and channel attention for VSR.

Conclusions:

• The proposed SAA network effectively enhances video super-resolution.
• Integrating attention-aware mechanisms in intermediate features boosts model performance.
• The C3AM module contributes to improved feature representation and VSR quality.