Extended Universal Joint Source-Channel Coding for Digital Semantic Communications: Improving Channel-Adaptability

Abstract

Recent advances in deep learning (DL)-based joint source-channel coding (JSCC) have enabled efficient semantic communication in dynamic wireless environments. Among these approaches, vector quantization (VQ)-based JSCC effectively maps high-dimensional semantic feature vectors into compact codeword indices for digital modulation. However, existing methods, including universal JSCC (uJSCC), rely on fixed, modulation-specific encoders, decoders, and codebooks, limiting adaptability to fine-grained SNR variations. We propose an extended universal JSCC (euJSCC) framework that achieves SNR- and modulation-adaptive transmission within a single model. euJSCC employs a hypernetwork-based normalization layer for fine-grained feature vector normalization and a dynamic codebook generation (DCG) network that refines modulation-specific base codebooks according to block-wise SNR. To handle block fading channels, which consist of multiple coherence blocks, an inner-outer encoder-decoder architecture is adopted, where the outer encoder and decoder capture long-term channel statistics, and the inner encoder and decoder refine feature vectors to align with block-wise codebooks. A two-phase training strategy, i.e., pretraining on AWGN channels followed by finetuning on block fading channels, ensures stable convergence. Experiments on image transmission demonstrate that euJSCC consistently outperforms state-of-the-art channel-adaptive digital JSCC schemes under both block fading and AWGN channels.

Figures (10)

• Figure 1: System overview of euJSCC, which performs block-wise feature vector encoding-decoding, VQ, and modulation.
• Figure 2: An illustration of slimmable operation on (a) convolutional layer, (b) transposed convolutional layer, and (c) GDN layer.
• Figure 3: The overall euJSCC encoder–decoder architecture, decomposed into inner and outer modules. "CNN", “T-CNN”, and “Slimmable T-Convolutional Layer” denote CNN blocks, transposed CNN blocks, and slimmable transposed convolutional layer, respectively. The blue-highlighted path shows an example processing flow for the $i$-th coherence block.
• Figure 4: An illustration of the proposed DCG network, which dynamically generates an SNR-adaptive codebook conditioned on $\eta$. For each modulation order $k$, a base codebook $\bar{\mathbf{C}}_{k}$, an intra-codeword adaptation $\mathtt{MLP}_{\boldsymbol{\psi}_{\mathsf{intra}, k^{(i)}}}$, and an inter-codeword adaptation $\mathtt{MLP}_{\boldsymbol{\psi}_{\mathsf{inter}, k^{(i)}}}$ are separately maintained, and the components corresponding to the given $k=k^{(i)}$ according to $\eta^{(i)}$ are activated. Both $\mathtt{MLP}_{\boldsymbol{\psi}_{\mathsf{intra}, k^{(i)}}}$ and $\mathtt{MLP}_{\boldsymbol{\psi}_{\mathsf{inter}, k^{(i)}}}$ follow the structure $\mathtt{MLP}_{\boldsymbol{\psi}_{k^{(i)}}}$ shown on (b), where $D_{\mathtt{MLP}}=D_{k^{(i)}}$ and $D_{\mathtt{MLP}}=m_{k^{(i)}}$, respectively.
• Figure 5: (a) Adaptation process for (I)GDN parameters of $\mathtt{HN\text{–}(I)G}$ layer. The dashed arrow from $\eta^{(i)}$ to $k^{(i)}$ indicates that $k^{(i)}$ is determined according to $\eta^{(i)}$. (b) Network architecture of hypernetwork $h_{\boldsymbol{\xi}_{\mathsf{(I)GDN}}}$ in $\mathtt{HN\text{–}(I)G}$ layer.

Theorems & Definitions (4)

• Remark 1: Standard GDN and inverse GDN
• Remark 2: Case of $N \neq TU$
• Remark 3: $\mathtt{HN\text{-}(I)G}$ layers in $\mathcal{F}_{\boldsymbol{\Theta}}$ and $\mathcal{G}_{\boldsymbol{\Phi}}$
• Remark 4: $\mathtt{HN\text{-}L}$ layer in $\mathcal{C}_{\boldsymbol{\Psi}}$