EUVIMEDEUVIMED entdecken
Crossref

A Lightweight Gated Convolution and Attention Joint Source–Channel Coding Architecture for Bandwidth-Limited Wireless Image Transmission

Helin Yang, Junhong Zhang, Changyuan Xu, Zeqi Huang, Jiawen Kang, Jiangtian Nie

IEEE Transactions on Wireless Communications · 2026 · 25 · 19193-19208

DOI: 10.1109/twc.2026.3706543 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Wireless Communications, 2026. Verfasst von Helin Yang und 5 weiteren Autor:innen. Die Arbeit nennt 46 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Fujian Provincial Natural Science Foundation of China, Research Funds for the Central Universities.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Deep learning-based joint source and channel coding (JSCC) is a good alternative to separation-based schemes for wireless image transmission, especially when bandwidth is tight and the signal-to-noise ratio (SNR) is low. However, most learned designs use one fixed backbone. Convolutional neural networks (CNNs) work well at low SNR, but they have limited global context. Transformers can model long-range structure, but they can be fragile and expensive. To this end, we propose Gated Mixture of Convolution and Attention based JSCC (GMCA-JSCC), an adaptive and lightweight architecture that blends convolution and attention using an SNR-driven gate. The attention path employs channel-adaptive deformable stripe attention (CADSA), and the model applies multi-level SNR conditioning across structural routing, feature-wise linear modulation, and attention hyperparameters, enabling smooth adaptation from low to high SNR. Evaluated across standard datasets and over additive white Gaussian noise (AWGN) and fading channels, GMCA-JSCC achieves strong fidelity and perceptual quality with graceful rate–SNR–distortion trade-offs. Under AWGN on Kodak, Bjøntegaard delta rate (BD-rate) results quantify substantial bandwidth savings at matched PSNR, reaching 44.5% relative to DynamicJSCC and 48.7% relative to SwinJSCC w/SNR-adaptive (SA). We further conduct robustness studies under more practical channel and receiver conditions. Ablations confirm the complementary roles of convolution and attention and the benefit of explicit SNR adaptation, while parameter count, floating-point operations (FLOPs), latency, and peak GPU memory remain practical for edge deployment.

Förderung

  • National Natural Science Foundation of China
  • Fujian Provincial Natural Science Foundation of China
  • Research Funds for the Central Universities
  • Xiaomi Young Talents Program

Zitieren

APA

Yang, H., Zhang, J., Xu, C., Huang, Z., Kang, J., & Nie, J. (2026). A Lightweight Gated Convolution and Attention Joint Source–Channel Coding Architecture for Bandwidth-Limited Wireless Image Transmission. IEEE Transactions on Wireless Communications, 25, 19193-19208. https://doi.org/10.1109/twc.2026.3706543

Vancouver

Yang H, Zhang J, Xu C, Huang Z, Kang J, Nie J. A Lightweight Gated Convolution and Attention Joint Source–Channel Coding Architecture for Bandwidth-Limited Wireless Image Transmission. IEEE Transactions on Wireless Communications. 2026;25:19193-19208. doi:10.1109/twc.2026.3706543

BibTeX
@article{Yang2026,
  author = {Yang, Helin and Zhang, Junhong and Xu, Changyuan and Huang, Zeqi and Kang, Jiawen and Nie, Jiangtian},
  title = {A Lightweight Gated Convolution and Attention Joint Source–Channel Coding Architecture for Bandwidth-Limited Wireless Image Transmission},
  journal = {IEEE Transactions on Wireless Communications},
  year = {2026},
  volume = {25},
  pages = {19193--19208},
  doi = {10.1109/twc.2026.3706543}
}
Kostenlos starten →