How to Adapt Wireless DJSCC Symbols to Rate Constrained Wired Networks?

TL;DR

This work addresses the end-to-end efficiency of deep joint source-channel coding (DJSCC) in hybrid wireless-wired networks by targeting two bottlenecks: redundancy in wireless-transmitted DJSCC symbols and the need to adapt to varying wired rate constraints. It introduces Rate-Controllable Wired Adaptor (RCWA), a modular framework that performs redundancy removal, continuous variable-rate coding via a Lagrangian multiplier, and rate-controlled encoding, while remaining compatible with existing DJSCC architectures. The approach leverages a Lagrangian RD objective, neural entropy modeling with hyperpriors, and context-aware high-resolution coding to achieve continuous rate control and end-to-end distortion reduction, validated on CIFAR-10 and high-resolution datasets like COCO/CLIC. Results show PSNR gains up to $0.7$ dB versus neural baselines and up to $4$ dB against digital baselines on CIFAR-10, along with robust performance under AWGN and Rayleigh fading, demonstrating practical impact for wired-core network resource utilization in next-generation networks.

Abstract

Deep joint source-channel coding (DJSCC) has emerged as a robust alternative to traditional separate coding for communications through wireless channels. Existing DJSCC approaches focus primarily on point-to-point wireless communication scenarios, while neglecting end-to-end communication efficiency in hybrid wireless-wired networks such as 5G and 6G communication systems. Considerable redundancy in DJSCC symbols against wireless channels becomes inefficient for long-distance wired transmission. Furthermore, DJSCC symbols must adapt to the varying transmission rate of the wired network to avoid congestion. In this paper, we propose a novel framework designed for efficient wired transmission of DJSCC symbols within hybrid wireless-wired networks, namely Rate-Controllable Wired Adaptor (RCWA). RCWA achieves redundancy-aware coding for DJSCC symbols to improve transmission efficiency, which removes considerable redundancy present in DJSCC symbols for wireless channels and encodes only source-relevant information into bits. Moreover, we leverage the Lagrangian multiplier method to achieve controllable and continuous variable-rate coding, which can encode given features into expected rates, thereby minimizing end-to-end distortion while satisfying given constraints. Extensive experiments on diverse datasets demonstrate the superior RD performance and robustness of RCWA compared to existing baselines, validating its potential for wired resource utilization in hybrid transmission scenarios. Specifically, our method can obtain peak signal-to-noise ratio gain of up to 0.7dB and 4dB compared to neural network-based methods and digital baselines on CIFAR-10 dataset, respectively.

Figures (13)

• Figure 1: Hybrid transmission with dense user access, which leads to transmission bottleneck in the core network.
• Figure 2: A hybrid transmission scheme RCWA-DJSCC for E2E communications, where the proposed RCWA including a redundancy eliminator, a Lagrangian multipler predictor, a feature adaptor and a feature encoder-decoder pair is integrated for efficient and reliable wired transmission. All modules in the scheme are realized by DNNs.
• Figure 3: Neural image compression frameworks. (a) Compression with hyperprior $\boldsymbol{z}$. (b) The extended compression framework with gain units for variable-rate coding.
• Figure 4: Network structures of redundancy eliminator and reconstruction module. IGDN denotes a generalized normalization layer proposed in ballé2016 for image compression. RB($M,M$) denotes residual block with $M$ input channels and $M$ output channels. AF($M$) denotes attention module proposed in Xu with intermediate features of $M$ channels. UB($M$) denotes upsample block with $M$ output channels. Basically, $M$ is set to 256 in this paper.
• Figure 5: Attention enabled variable-rate feature coding. RB and UB are the residual block and the upsampling block illustrated in Fig. \ref{['Fig:RE_RM']}. CB($M$,$M$) denotes the coding block in (a) with $M$ input channels and $M$ output channels. DB($M$,$N$) denotes downsampling block with $M$ input channels and $N$ output channels. Basically, $M$ and $N$ are set to 256 and 192, respectively.

Theorems & Definitions (4)

• Remark 1
• Remark 2
• Remark 3
• Remark 4