• Semantic communications transmit only task-relevant information, reducing bandwidth for interactive CV. • Editable-DeepSC introduces joint editing-channel coding (JECC) to preserve semantic mutual information. • Uses StyleGAN inversion to compactly encode high-dimensional facial attributes. • SNR-aware channel coding fine-tunes models to adapt to dynamic noise conditions. • Experiments show superior editing quality and significant bandwidth savings, even at high resolution. • Demonstrates robustness under out-of-distribution settings, maintaining performance across varied data.
Article Summaries:
- Editable‑DeepSC introduces a cross‑modal semantic communication framework for facial editing, addressing the mismatch between conventional data‑centric networks and interactive computer‑vision tasks. The authors argue that separate transmission and editing pipelines lose critical semantic information, proposing a Joint Editing‑Channel Coding (JECC) scheme that iteratively matches editing attributes within the communication chain. High‑dimensional facial data are compactly encoded using StyleGAN inversion, while a signal‑to‑noise ratio (SNR)‑aware channel coder adapts to dynamic noise conditions via model fine‑tuning. Experiments show the method delivers superior editing quality, reduces bandwidth usage, and remains robust under high‑resolution and out‑of‑distribution scenarios.
Sources: