.png){kind=icon}
Return
Dual-contextual scalable video coding for human-machine collaborative intelligence
.png){kind=icon}
PRE
AI
DOI：10.1016/j.displa.2026.103638.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Generating in nearly every aspect of daily life, video is driving the flourishing development of intelligent visual analysis. With its explosive volume growth, efficient and reliable human-machine collaborative video coding systems are increasingly needed. However, state-of-the-art human-oriented compression methods still rely on resource-intensive reconstruction of the full pixel space before machine analysis, whereas machine-oriented compression methods that use bitstream directly for analysis often struggle to reconstruct high quality video. To address these under-explored challenges, we propose a novel scalable video compression framework that balances human and machine vision by progressively compressing decoupled contextual. Specifically, we employ a sparse feature extractor to construct a sparse compression loop with task-specific transformations to effectively serve various machine analysis tasks. Based on decoupled non-overlapping subsets, we integrate a prediction decoupling module and a cross-layer entropy model into an enhancement compression loop to balance the reconstruction quality and coding efficiency. Extensive experimental results demonstrate that the proposed method outperforms competing methods on several machine analysis tasks while achieving comparable or even better reconstruction quality.
Keywords:
Video coding
Human visual
Machine visual
Scalable coding
Journal
IF:
3.4
Papers:
2.3K
Citations:
3.2K
Organization
Cited Papers
No cited papers available