MDEformer: Mixed Difference Equation Inspired Transformer for Compressed Video Quality Enhancement.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning excels in video quality enhancement but often relies on manual network design.
• Existing methods underutilize multiscale feature similarity and overlook partition boundary impacts in compressed videos.

Purpose of the Study:

• To propose a novel, interpretable transformer network for compressed video quality enhancement.
• To provide a principled approach for network design guided by mixed difference equations.

Main Methods:

• Developed the Mixed Difference Equation inspired Transformer (MDEformer).
• Incorporated cross-layer cross-attention aggregation (CCA) modules for multiscale artifact similarity.
• Integrated partition boundary smoothing (PBS) modules to mitigate boundary artifacts.

Main Results:

• MDEformer effectively removes compression artifacts and recovers texture details.
• The CCA modules exploit multiscale similarity for artifact removal.
• PBS modules significantly reduce partition boundary impact on video quality.

Conclusions:

• MDEformer offers a principled and interpretable approach to compressed video quality enhancement.
• The proposed method surpasses state-of-the-art techniques in objective and visual quality metrics.