RRDnCNN

RRDnCNN applies a degradation-aware deep neural network to jointly restore and reconstruct compressed, down-sampled video by removing compression artifacts and performing super-resolution.

Key Features:

• Integration of Deep Learning: Leverages deep learning methods to enhance the quality of videos produced by down-sampling-based compression frameworks.
• Artifact Removal and Super-Resolution: Performs simultaneous artifact removal and super-resolution to address degradation introduced by compression and sub-sampling.
• Degradation-Aware Technique: Mitigates adverse effects of compression and sub-sampling across configurations including Random Access, Low Delay P, and All Intra.
• End-to-End Network Architecture: Uses a u-shaped design with up-sampling skip connections to improve learning, enable reconstruction to use restoration features, and alleviate gradient vanishing in deep networks.
• Performance Improvements: RR-DnCNN v2.0 achieves a 17.02% BD-rate reduction on UHD resolution videos anchored by the H.265/HEVC codec.

Scientific Applications:

• Medical imaging: Enhances resolution and reduces compression artifacts in medical video data.
• Surveillance: Improves visual quality of surveillance video for analysis and monitoring.
• Remote sensing: Enhances compressed remote sensing video by restoring details and reducing artifacts.
• Visual data integrity domains: Supports any domain where maintaining high visual data integrity after compression is crucial.

Methodology:

Uses a degradation-aware deep learning approach with a joint restoration-reconstruction end-to-end network; employs a u-shaped architecture with up-sampling skip connections to address gradient vanishing and targets down-sampling-based video compression frameworks with super-resolution as post-processing.