Lightweight Low-SNR-Robust Semantic Communication System for Autonomous Driving

Abstract: Image transmission for vehicle-to-vehicle collaborative perception in autonomous driving faces challenges including limited on-board terminal resources, time-varying wireless channel fading, and poor robustness under low signal-to-noise (SNR) ratio. Traditional separate source-channel coding schemes suffer from the cliff effect, while existing semantic communication models are limited by large parameter sizes and weak digital compatibility. This paper proposes a lightweight, low-SNR-robust deep joint source-channel coding (JSCC) semantic communication system. First, structured pruning is implemented based on batch normalization layer scaling factors and L1 regularization, which significantly reduces model complexity while ensuring image reconstruction quality. Second, a uniform quantization and M-QAM modulation scheme adapted to JSCC features is designed, and a training-deployment separation strategy is adopted to address the non-differentiable quantization problem, enabling compatibility with existing digital communication systems. Simulation results on the Cityscapes dataset show that the pruned model maintains comparable performance and robustness to the original one, even with over half of its parameters removed. Notably, the proposed scheme exhibits significant advantages over conventional communication methods under low SNR conditions.

• The paper introduces a lightweight V2V semantic communication system that combines BN-based structured pruning with quantized M-QAM transmission, reducing parameters from 6.25M to 4.02M at 50% pruning while decreasing PSNR by only 0.43 dB at 25 dB SNR.
• The proposed digital interface preserves the low-SNR resilience of deep JSCC, with 256QAM staying within 0.5 dB of analog transmission above 15 dB SNR and all tested modulation orders maintaining SSIM above 0.7 in extremely noisy conditions.
• The system outperforms conventional BPG-LDPC coding below roughly 14.5 dB SNR and shows that structured pruning can preserve representational capacity more effectively than reducing the bandwidth compression ratio, although real-world vehicular validation remains necessary.

Motivation and problem statement

Vehicle-to-vehicle (V2V) collaborative perception requires transmitting camera images between vehicles under time-varying Rayleigh fading, limited bandwidth, and severely constrained on-board compute and storage. Conventional separate source-channel coding (e.g., BPG image coding plus LDPC channel coding with QAM modulation) suffers from the cliff effect: below a channel-quality threshold, LDPC decoding fails catastrophically and reconstruction quality collapses. Existing deep joint source-channel coding (JSCC) approaches avoid this failure mode but present two deployment obstacles: their convolutional encoder–decoder architectures are too large for vehicular edge platforms, and they typically transmit analog real-valued symbols that are incompatible with existing digital communication infrastructure (2604.20278).

The paper addresses both obstacles simultaneously. Its contributions are a structured pruning pipeline based on batch normalization (BN) scaling factors with L1 regularization, and a uniform quantization plus M-QAM modulation scheme adapted to JSCC features using a training–deployment separation strategy to sidestep the non-differentiability of quantization.

System model

The system considers a single-antenna V2V link over a slow Rayleigh fading channel with additive white Gaussian noise, where the received signal is $$\mathbf{y} = \mathbf{h} \odot \mathbf{s} + \mathbf{n}$$ subject to an average power constraint. A convolutional JSCC encoder maps an image $\mathbf{x} \in \mathbb{R}^n$ to semantic features $\mathbf{z} \in \mathbb{R}^k$ with bandwidth compression ratio $k/n = 2/3$; a Sigmoid output layer constrains $\mathbf{z}$ to $(0,1)$, which is what makes uniform quantization natural. The receiver performs minimum-distance demodulation and reconstructs the image via the JSCC decoder. Evaluation uses PSNR and SSIM on Cityscapes street-scene images resized to $512 \times 512$.

Structured pruning for lightweight deployment

The pruning approach follows network-slimming principles: L1 regularization is applied to BN scaling factors $\eta$ during sparse training, driving unimportant channels toward zero, after which channels with the smallest $|\eta|$ values are globally removed at rate $\gamma$ and the surviving model is fine-tuned with MSE loss alone. Because pruning is structured rather than unstructured, the resulting network retains regular tensor shapes and can be accelerated on commodity hardware without specialized sparse-computation support—an important practical consideration for vehicle-mounted inference.

The empirical results support the central lightweighting claim. At $\mathbf{x} \in \mathbb{R}^n$0, parameters drop from 6.25 M to 4.02 M and MACs from 20.24 G to 18.03 G, yet PSNR at SNR = 25 dB falls only from 31.42 dB to 30.99 dB with SSIM unchanged at 0.91. Even at $\mathbf{x} \in \mathbb{R}^n$1 (3.12 M parameters), PSNR remains 30.36 dB. Only aggressive pruning at $\mathbf{x} \in \mathbb{R}^n$2 degrades noticeably (28.57 dB PSNR, 0.87 SSIM). Notably, the lightly pruned model ($\mathbf{x} \in \mathbb{R}^n$3) slightly exceeds the unpruned baseline in both metrics, which the authors attribute to a regularization effect improving generalization—a claim consistent with, though not separately ablated in, the reported results.

A particularly informative comparison pits the pruned model ($\mathbf{x} \in \mathbb{R}^n$4, $\mathbf{x} \in \mathbb{R}^n$5) against a model with a directly reduced bandwidth compression ratio ($\mathbf{x} \in \mathbb{R}^n$6) of similar parameter count. The pruned model outperforms the bandwidth-reduced model across the SNR range, indicating that removing redundant channels preserves representational capacity more effectively than reducing transmitted symbol count. This is a useful design insight: for fixed on-board memory budgets, structural sparsity is preferable to bandwidth reduction.

Low-SNR-robust digital transmission

The quantization-modulation scheme exploits the Sigmoid-bounded encoder output: each feature element is uniformly quantized to one of $\mathbf{x} \in \mathbb{R}^n$7 levels on $\mathbf{x} \in \mathbb{R}^n$8 and mapped deterministically to an M-QAM constellation point. The floor operation underlying quantization is non-differentiable, so the authors adopt a training-deployment separation strategy: during training, unquantized features pass through an equivalent noisy analog channel so gradients flow normally; at inference, quantization, constellation mapping, hard demodulation, and equalization are inserted without retraining. The decoder's exposure to noise during training confers robustness to both quantization error $\mathbf{x} \in \mathbb{R}^n$9 and demodulation error $\mathbf{z} \in \mathbb{R}^k$0, provided $\mathbf{z} \in \mathbb{R}^k$1, i.e., the quantization is fine relative to the equivalent training noise.

The results bear out this analysis. With 256QAM, the PSNR gap to ideal analog transmission is less than 0.5 dB above SNR = 15 dB, and even 4QAM sustains PSNR above 20.4 dB at SNR = 15 dB. Under extremely low SNR, all modulation orders retain SSIM above 0.7 while the BPG-LDPC baseline has already collapsed to SSIM = 0. Qualitatively, the digital scheme at SNR = 1 dB achieves 27.76 dB PSNR and 0.90 SSIM—visually acceptable output where the BPG scheme produces severe block artifacts—and the reconstructed images appear cleaner than the analog JSCC outputs, which the authors attribute to hard-decision demodulation suppressing small noise perturbations. This denoising-by-discretization observation is one of the paper's more interesting findings, suggesting that constellation discretization can act as a mild implicit error-correction mechanism for learned decoders.

Against the separate-coding baseline, the crossover behavior is as expected: BPG-LDPC wins above roughly 14.5 dB in PSNR (35.55 dB / 0.95 SSIM at high SNR versus ~31 dB / 0.91 for JSCC), reflecting the asymptotic optimality of separation under good channel conditions, while deep JSCC dominates decisively below that threshold because it never experiences decoding failure. For autonomous driving, where link quality is frequently poor, the low-SNR regime is arguably the operationally relevant one.

Limitations and open questions

Several caveats qualify these results. First, all evaluation is simulation-based on Cityscapes with slow Rayleigh fading; no over-the-air experiments, Doppler effects, or fast-fading channels are considered, so the claimed vehicular suitability rests on simulated channel conditions. Second, the training SNR is fixed at 25 dB, and the paper does not evaluate performance mismatch when deployment SNRs differ substantially from the training condition or explore SNR-adaptive variants. Third, the comparison baseline is a single conventional configuration (BPG + rate-3/4 LDPC + QAM); stronger baselines such as adaptive-rate JSCC or learned digital schemes with soft demodulation are not included. Fourth, although the paper argues latency benefits qualitatively through reduced MACs, no end-to-end latency measurements or actual embedded-platform benchmarks are reported—the resource-constraint motivation is addressed only via parameter/MAC counts. Finally, the mechanism behind the apparent generalization benefit of light pruning and the denoising effect of hard decisions is hypothesized rather than analyzed, leaving open why discretization helps the learned decoder and whether this holds across channel models and datasets.

Conclusion

This paper presents a coherent engineering solution combining BN-based structured pruning with a train-deploy-separated uniform quantization and M-QAM interface for deep JSCC in V2V collaborative perception. The quantitative evidence shows that half the model can be removed with negligible reconstruction loss, that 256QAM digital transmission closely tracks analog JSCC performance, and that the system maintains usable image quality at SNRs where separate coding fails entirely. The main open issues are validation under realistic vehicular channel dynamics and hardware, robustness to train/deploy SNR mismatch, and a mechanistic account of the observed discretization-induced denoising.