New PVSC Framework Slashes Video Bandwidth by 75% While Preserving Perception

Electrical Engineering researchers Yinhuan Huang and Zhijin Qin have introduced PVSC (Perception-Aware Video Semantic Communication), a novel framework that rethinks video transmission for bandwidth-limited, latency-constrained wireless links. Traditional separated source-channel systems optimize for bit-level reliability, often failing under short-blocklength transmission, while pixel-level distortion metrics don't align with human perception. PVSC addresses these issues by eliminating explicit motion-vector transmission and instead leveraging spatio-temporal feature coding to generate compact, channel-robust symbol streams. The framework also specifies side-information formatting, reference-buffer management, and lightweight rate control, enabling stable receiver-side reconstruction and bandwidth-adaptive inference with a single model.

In extensive experiments across diverse datasets, resolutions, and channel conditions, PVSC demonstrated dramatic bandwidth savings without compromising perceptual quality. Compared to the engineered VTM + 5G LDPC baseline, PVSC saves up to about 75% bandwidth at comparable LPIPS (Learned Perceptual Image Patch Similarity) and 87% at DISTS (Deep Image Structure and Texture Similarity). Critically, the system runs real-time inference on a single NVIDIA RTX 4090 GPU, making it practical for deployment. This work points toward a future where ultra-high-resolution streaming and immersive services can be delivered over constrained wireless links with human-centric quality.