.png){kind=icon}
.png){kind=icon}
Return
Adaptive joint source-channel coding with regulated autoencoder
.png){kind=icon}
OA
AI
DOI：10.1016/j.dcan.2026.03.008.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
This paper introduces ReJSCC, a unified deep Joint Source-Channel Coding (JSCC) framework designed for cross-modal adaptability to dynamic channel conditions. Unlike prior methods that rely on context-aware fusion of image features and channel SNR, ReJSCC employs systematic SNR-conditioned regulation to enable a single model to guide encoding and decoding processes across varying noise levels. The proposed framework integrates a regulating module that transforms encoder features through channel-specific scaling derived solely from SNR, paired with a symmetric deregulating module to steer channel-aware decoding. Extensive experiments demonstrate ReJSCC’s consistent superiority for both image and speech transmission, outperforming conventional JSCC and its robust version across broad SNR ranges. Ablation studies reveal that intermediate feature modulation during encoding/decoding surpasses pre-post channel intervention, while explicit SNR guidance to the encoder enhances reconstruction fidelity. Notably, ReJSCC shows strong robustness to inaccurate SNR estimation, demonstrating its practical resilience in real- world environments. Surprisingly, experiments suggest that integrating image feature maps for SNR adaptation may be nonessential, simplifying architectural design.
Keywords:
Joint source channel coding
Image transmission
Speech transmission
Regulating network
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
7.5
Papers:
406
Citations:
3.5K