Hai-Long Qin¹, Jincheng Dai¹†, Sixian Wang¹, Xiaoqi Qin¹,
Shuo Shao², Kai Niu¹, Wenjun Xu¹, Ping Zhang¹†
¹ Beijing University of Posts and Telecommunications (BUPT)
² University of Shanghai for Science and Technology (USST)
† Corresponding Authors
Accepted by IEEE Communications Standards Magazine (COMSTD), 2025
We establish a standardized coding workflow for semantic communication that moves beyond transmitting deep features and instead conveys compact, context-aware general semantic representations through:
- 📝 Tokenization: Breaking data into embedded tokens
- 🔄 Reorganization: Merging semantically similar tokens based on contextual similarity
- ⚡ Quantization: Optional mapping to discrete codes approaching "one code, one concept"
Our method uses just 10–30 semantic tokens to achieve better reconstruction quality (lower FID) and higher fidelity (higher PSNR) than traditional neural coding, especially at low data rates.
If you find this work useful, please cite:
@article{qin2025semcod,
author = {H. L. Qin and J. Dai and S. Wang and X. Qin and S. Shao and K. Niu and W. Xu and P. Zhang},
title = {Neural Coding Is Not Always Semantic: Toward the Standardized Coding Workflow in Semantic Communications},
journal = {IEEE Commun. Stand. Mag.},
volume = {9},
number = {4},
pages = {24--33},
year = {2025}
}