Collaborative Feedback Discriminative Propagation for Video Super-Resolution.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Existing video super-resolution (VSR) methods rely on spatial and temporal information, often using temporal propagation with alignment.
• Inaccurate alignment in VSR can cause artifacts, accumulating during propagation and degrading video restoration quality.
• Simple forward or backward propagation struggles with complex motion and occlusions in videos.

Purpose of the Study:

• To propose a novel Collaborative Feedback Discriminative (CFD) method for video super-resolution.
• To address limitations of inaccurate alignment and improve spatial-temporal information modeling in VSR.
• To enhance the performance and robustness of existing VSR models.

Main Methods:

• Developed a Discriminative Alignment Correction (DAC) method to mitigate artifacts from inaccurate alignment.
• Proposed a Collaborative Feedback Propagation (CFP) module utilizing feedback and gating for simultaneous forward/backward temporal feature exploration.
• Integrated DAC and CFP modules into standard VSR networks for validation.

Main Results:

• The proposed CFD method significantly reduces artifacts caused by inaccurate feature alignment.
• CFP module effectively models complex spatial and temporal information across different timesteps.
• Experimental results show improved performance of VSR models integrated with DAC and CFP.

Conclusions:

• The CFD method offers an effective solution for improving video super-resolution.
• The approach enhances VSR by correcting alignment inaccuracies and better utilizing spatio-temporal data.
• CFD improves existing VSR models' effectiveness while maintaining lower computational complexity.