A Scene-Adaptive Super-Resolution Framework for Video Compression.

Area of Science:

• Computer Science
• Signal Processing
• Machine Learning

Background:

• Video compression is crucial for efficient large-scale video delivery, impacting bandwidth and storage costs.
• Super-resolution (SR) models enhance low-resolution video streams, but fixed temporal updates in SR-assisted compression can be inefficient.
• Inefficient model updates lead to wasted bitrate when not aligned with actual scene changes.

Purpose of the Study:

• To develop a scene-adaptive super-resolution video compression framework (SASVC) for improved efficiency.
• To reduce bitrate overhead by updating reconstruction models only when scene changes are detected.
• To maintain scene-specific reconstruction fidelity and practical decoder-side deployment.

Main Methods:

• SASVC detects scene changes using frame-wise grayscale differences.
• Updates are applied only to compact adapter modules during detected content transitions.
• Model updates are compressed using chained differencing, quantization, and entropy coding.

Main Results:

• SASVC consistently outperforms SRVC-style baselines and conventional codecs in BD-rate/PSNR.
• Significant BD-rate gains observed: -41.33% (Vimeo), -53.49% (Xiph), -51.53% (UVG), -39.83% (MCL-JCV).
• Real-time 1080p reconstruction at 125 FPS achieved on an NVIDIA RTX 3080 Ti GPU.

Conclusions:

• Scene-adaptive model updates in SASVC significantly improve video compression efficiency.
• The framework reduces unnecessary model-stream overhead while preserving reconstruction quality.
• SASVC offers practical decoder-side deployment with real-time performance.