Learning a Deep Dual Attention Network for Video Super-Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning video super-resolution (SR) methods often use convolutional neural networks (CNNs) with motion compensation.
• Existing methods suffer from inaccurate motion estimation due to downscaling and treat intermediate features uniformly, hindering high-frequency detail recovery.

Purpose of the Study:

• To propose a novel deep dual attention network (DDAN) for accurate video super-resolution.
• To address limitations in motion estimation accuracy and feature utilization in previous SR methods.

Main Methods:

• The proposed DDAN comprises a motion compensation network (MCNet) and a super-resolution reconstruction network (ReconNet).
• MCNet uses a pyramid approach for optical flow and incorporates detail components from LR frames to reduce mis-registration errors.
• ReconNet employs dual attention mechanisms within residual units (residual attention units) to focus on informative features for detail enhancement.

Main Results:

• The DDAN effectively exploits spatio-temporal features for precise video super-resolution.
• Experimental results on multiple datasets show superior quantitative and qualitative performance compared to state-of-the-art methods.
• The method demonstrates significant improvements in recovering high-frequency details.

Conclusions:

• The proposed deep dual attention network (DDAN) offers a robust solution for video super-resolution.
• By enhancing motion compensation and feature attention, the method achieves state-of-the-art performance.
• This work contributes to advancing the accuracy and detail recovery capabilities in video SR.