Human Visual Perception Based Edge‑Cloud Super‑Resolution Framework and RedVQA Quality Assessment at XiaoHongShu

XiaoHongShu’s video business is growing rapidly, prompting the need to improve user‑perceived video quality while reducing bandwidth costs. At the LiveVideoStackCon 2023 Shanghai conference, Jianhan, head of video‑image algorithm research, presented an innovative edge‑cloud combined super‑resolution framework driven by human visual perception.

The talk covered four main parts:

1. An overview of XiaoHongShu’s video processing architecture, which spans production, cloud, and consumption sides. The architecture aims to simultaneously enhance visual quality and lower bandwidth consumption.

2. Introduction of RedVQA, a self‑developed AI‑based no‑reference video quality assessment (VQA) model that aligns with human visual perception. RedVQA can evaluate video quality at any point in the processing pipeline and guide both quality‑enhancement and encoding optimizations.

3. Design of an edge‑cloud combined super‑resolution solution. The cloud generates a dedicated super‑resolution bitrate tier, while the client runs a lightweight super‑resolution algorithm after decoding. Bandwidth‑peak prediction and quality‑vs‑bitrate trade‑off analysis (using RedVQA) decide when and how to enable the super‑resolution tier.

4. Summary and outlook, emphasizing the importance of intelligent quality assessment, global system optimization, and deeper edge‑cloud collaboration for future “narrow‑band HD” video delivery.

Key technical insights included:

Module‑level optimization through newer codecs (H.265, AV1, H.266) and their deployment challenges.

Cross‑technology fusion by combining content‑aware encoding, scene classification, and AI‑driven quality enhancement.

Global system optimization that balances QoS (e.g., latency, stall rate) and QoE (user‑perceived quality) while achieving significant bandwidth savings.

Practical performance targets for edge‑side super‑resolution (≤10 ms latency, ≤100 mAh power, ≤GFLOPS compute).

AB experiments on iPhone XR/XS devices demonstrated positive QoS improvements and notable bandwidth reduction.

The presentation concluded with a vision for more intelligent, fine‑grained quality assessment and continued optimization of cloud‑side “narrow‑band HD” transcoding.