RR-DnCNN v2.0: Enhanced Restoration-Reconstruction Deep Neural Network for Down-Sampling-Based Video Coding.

Area of Science:

• Computer Vision
• Deep Learning
• Video Compression

Background:

• Deep learning enhances video coding, particularly super-resolution for down-sampled videos.
• Compression artifacts and degradation complicate super-resolution tasks.
• Integrating artifact removal before super-resolution can remove useful features.

Purpose of the Study:

• To develop an end-to-end deep neural network for video restoration and reconstruction.
• To address challenges posed by compression artifacts and sub-sampling in super-resolution.
• To improve video quality and coding efficiency in video compression frameworks.

Main Methods:

• Proposed an end-to-end restoration-reconstruction deep neural network (RR-DnCNN) utilizing degradation-aware techniques.
• Investigated the efficacy of Random Access configuration degradation for training across various compression types.
• Redesigned the network architecture to a u-shaped network (RR-DnCNN v2.0) with up-sampling skip connections to improve feature leveraging.

Main Results:

• Demonstrated that Random Access degradation covers other types like Low Delay P and All Intra for training.
• The redesigned RR-DnCNN v2.0 network overcomes gradient vanishing issues present in straightforward chained networks.
• Achieved a 17.02% BD-rate reduction on UHD resolution for all-intra, outperforming previous methods.

Conclusions:

• The proposed RR-DnCNN v2.0 effectively handles video degradation and compression artifacts for super-resolution.
• The degradation-aware technique and novel architecture significantly improve video restoration and reconstruction.
• This approach offers substantial coding efficiency gains for high-resolution video compression.