Selective Depthwise Separable Convolution for Lightweight Joint Source-Channel Coding in Wireless Image Transmission

Abstract: Depthwise separable convolutional (DSConv) layers have been successfully applied to deep learning (DL)-based joint source-channel coding (JSCC) schemes to reduce computational complexity. However, a systematic investigation of the layerwise and ratio-wise replacement of standard convolutional (Conv) layers with DSConv layers in JSCC systems for wireless image transmission remains largely unexplored. In this letter, we propose a configurable lightweight JSCC framework that incorporates a selective replacement strategy, enabling flexible substitution of standard Conv layers with DSConv layers at various layer positions and replacement ratios. By adjusting the proportion of layers replaced, we achieve different model compression levels and analyze their impact on reconstruction performance. Furthermore, we investigate how replacements at different encoder and decoder depths influence reconstruction quality under a fixed replacement ratio. Our results show that Conv-to-DSConv replacement at intermediate layers achieves a favorable complexity-performance trade-off, revealing layer-wise redundancy in DL-based JSCC systems. Extensive experiments further demonstrate that the proposed framework achieves substantial parameter reduction with only slight performance degradation, enabling flexible complexity-performance trade-offs for resource-constrained edge devices.

• The paper introduces DSC-JSCC, which selectively replaces standard convolutions with depthwise separable variants to significantly reduce computational complexity in wireless image transmission.
• It demonstrates up to an 82.3% reduction in parameters and optimal balance between PSNR and LPIPS through intermediate-layer replacements, ensuring high image reconstruction quality.
• The approach offers granular control over replacement ratios and positions, enabling efficient deployment on resource-constrained edge devices and scalability to multi-modal semantic communications.

Selective Depthwise Separable Convolution for Lightweight JSCC in Wireless Image Transmission

Motivation and Context

Modern edge devices face stringent restrictions on computational, memory, and energy resources, limiting the practical deployment of deep learning-based joint source-channel coding (DL-based JSCC) for wireless image transmission. Recent advances in semantic communication systems for 6G have leveraged DL-based JSCC to optimize the end-to-end mapping between source images and channel inputs, prioritizing semantic fidelity over traditional bit error rates. However, most JSCC frameworks are computationally intensive due to standard convolutional layers. Previous literature has introduced depthwise separable convolution (DSConv) for lightweight architectures, but systematic analysis of layer-wise and ratio-wise replacement strategies remains incomplete.

Approach and Architecture

The paper introduces DSC-JSCC, a configurable, lightweight JSCC framework that selectively replaces standard convolutional (Conv) and transposed convolution (TConv) layers with their depthwise separable counterparts (DSConv and DSTConv) in both encoder and decoder. The central innovation lies in the selective replacement strategy, enabling granular control over:

• Replacement ratios: Fractional substitutions of Conv/TConv layers (20%, 40%, 60%, 80%, 100%), allowing tunable model sparsity and compression.
• Replacement positions: Targeted substitution at early, intermediate, or late layers within encoder/decoder, exploiting layer-wise redundancy and computational load distribution.

Each DSConv layer decomposes standard convolution into a spatial depthwise operation followed by pointwise channel mixing. DSTConv similarly splits transposed convolution. This architectural modularity facilitates empirical analysis of complexity-performance trade-offs.

Experimental Analysis and Numerical Results

Evaluation leverages the CelebA-HQ dataset, with performance quantified using PSNR (pixel-level fidelity) and LPIPS (perceptual quality). The baseline Deep-JSCC model [r3] is compared against DSC-JSCC variants.

Key findings include:

• Complexity Reduction: DSC-JSCC-60 (E2D2) exhibits a dramatic reduction in parameters (down by 82.3% versus baseline) and FLOPs (down by 75.3%), illustrating the efficiency of intermediate-layer replacement.
• Trade-off Performance: DSC-JSCC-20 matches or outperforms Deep-JSCC at high SNR (e.g., PSNR at 19 dB), with minimal computational overhead. Higher ratios (≥80%) introduce noticeable performance degradation.
• Layer-wise Redundancy: Intermediate-layer replacements result in minimal loss in reconstruction quality, confirming that computational bottlenecks are concentrated mid-network.
• Position Effects: Analysis at a fixed 60% replacement ratio reveals that DSC-JSCC-60 (E3D2) achieves optimal PSNR/LPIPS, but DSC-JSCC-60 (E2D2) delivers near-equivalent quality with the lowest complexity, ideal for edge deployment.
• Low-SNR Behavior: Under low SNR (<5 dB), complexity-reduced models converge toward baseline performance, supporting the adoption of lightweight models in high-noise environments.

Strong numerical results: DSC-JSCC-60 (E2D2), with only 25.4K parameters and 205.9M FLOPs, achieves PSNR and LPIPS comparable to the Deep-JSCC model (143.7K parameters, 832.4M FLOPs), at significantly reduced computational cost.

Theoretical and Practical Implications

The selective replacement methodology exposes the redundancy in JSCC layer computation and enables quantifiable model compression limits without substantial performance loss. This aids in the theoretical understanding of layer-wise sensitivity in end-to-end semantic communication systems. In practical deployment, DSC-JSCC provides a blueprint for adaptive complexity-performance tuning, critical for real-time wireless image transmission on resource-constrained hardware.

The framework's modularity also enables extension to other semantic modalities (e.g., text, video) and adaptation to diverse channel models, provided differentiability is preserved. Control over replacement ratios and positions facilitates rapid architecture search—integration with NAS techniques could further automate optimal configuration selection.

Future Directions

Future research may address:

• Model generalization: Extension of selective DSConv replacement strategies to multi-modal semantic communication systems.
• Joint NAS optimization: Automated search for optimal replacement ratio/position using neural architecture search and reinforcement learning.
• Hardware integration: Deployment benchmarks on edge devices (e.g., mobile GPUs, FPGAs), quantifying real-world latency and energy savings.
• Robustness studies: Analysis of selective DSConv under adversarial conditions, non-Gaussian channels, or highly compressed regimes.

DSC-JSCC's principles could also catalyze new theoretical insights into the interplay between semantic fidelity and architectural sparsity in end-to-end communication systems.

Conclusion

The paper presents a systematic framework for selectively deploying depthwise separable convolution in DL-based JSCC, achieving substantial computational savings with minimal degradation in image reconstruction performance. The selective replacement strategy, analyzed across ratios and positions, uncovers layer-wise redundancy and informs design choices for practical wireless transmission in edge scenarios. Empirical results confirm effective complexity-performance trade-offs, positioning DSC-JSCC as a robust, flexible foundation for future semantic communication designs (2604.22338).