High-Frequency information prediction based Video Super Resolution

Advisor: Prof. Anurag Mittal, Computer Vision Lab, IIT Madras, India

Network Architecture

• Designed a novel 2-phase progressive-retrogressive training, and a dual motion warping frame alignment techiniques
• Designed a neural network for explicit refinement and fusion of high-frequency details of super resolved videos
• The system produced more visually appealing results than most SOTA methods with no noticeable temporal artifacts

Results

Ground Truth	Our Result

• This work was accepted at Winter Conference on Applications of Computer Vision (WACV 2020). The paper is available here