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

Jul 16, 2026

Depthwise separable convolutional (DSConv) layers reduce complexity in deep learning joint source-channel coding (JSCC), yet a systematic study of replacing standard convolution (Conv) with DSConv at different layer positions and replacement ratios for wireless image transmission has been missing. This letter presents a configurable lightweight JSCC framework with selective Conv-to-DSConv replacement, producing models with different complexities. It analyzes reconstruction performance effects and encoder/decoder depth impacts, showing favorable complexity-performance trade-offs, redundancy in intermediate layers, and substantial parameter reductions with slight degradation for edge devices.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages6

Reading Time15 min

Introduction

Semantic communication and DL-based JSCC

Motivation: edge constraints and need for lightweight models

Related lightweight JSCC methods

Proposed framework (DSC-JSCC)

Selective Depthwise Separable Convolution for Lightweight Joint Source-Channel Coding in Wireless Image Transmission Ming Ye, Member, IEEE, Kui Cai, Senior Member, IEEE, Cunhua Pan, Senior Member, IEEE, Zhen Mei, Member, IEEE, Wanting Yang, and Chunguo Li, Senior Member, IEEE arXiv :2604 .22338v2 [ ee ss .IV] 11 Jul 2026 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 layer-wise 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 Conv-to-DSConv replacement at different replacement ratiosand positions. By varying the replacement ratio, we obtain models with different computational complexities 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 atthe intermediate layers of the encoder and decoder 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. Index Terms—Lightweight joint source-channel coding, deep learning, depthwise separable convolution, wireless image transmission. I. INTRODUCTION SEMANTIC communication has emerged as a promising technology for sixth-generation (6G) communication systems, serving as an innovative paradigm for integrating communications and artiﬁcial intelligence [1] . Unlike traditional communication systems, it focuses on conveying the meaning of transmitted information instead of solely minimizing bit error rates. Deep learning (DL)-based joint source-channel coding (JSCC) has become a promising and widely adopted enabler of this paradigm [2], [3] . It directly maps the original information into continuous channel inputs, facilitating endto-end optimization of the communication system [4], [5] . M. Ye, K. Cai, and W. Yang are with the Singapore University of Technology and Design, Singapore 487372 (e-mail: [230198460@aa.seu.edu.cn](230198460@aa.seu.edu.cn); cai− [kui@sutd.edu.sg](kui@sutd.edu.sg); wanting− [yang@sutd.edu.sg](yang@sutd.edu.sg)). C. Pan and C. Li are with the National Mobile Communications Research Laboratory, Southeast University, Nanjing 210096, China (e-mail: [cpan@seu.edu.cn](cpan@seu.edu.cn); [chunguoli@seu.edu.cn](chunguoli@seu.edu.cn)). Z. Mei is with the School of Electronic and Optical Engineering, Nanjing University of Science and Technology, Nanjing 210094, China (e-mail: [meizhen@njust.edu.cn](meizhen@njust.edu.cn)). Recent advancements in DL, especially in image processing [6], [7] and natural language processing [8], have further accelerated the development of semantic communication. Nevertheless, the limited computational, memory, and energy resources of edge devices pose signiﬁcant challenges to the deployment of DL-based JSCC models. Reducing the model size of DLbased JSCC systems is therefore essential, as it directly affects real-time performance, storage requirements, energy consumption, and deployment ﬂexibility. Consequently, thereis a pressing need for lightweight DL-based JSCC methods that achieve acceptable performance with low computational complexity. Several recent studies have explored lightweight DL-based JSCC approaches [11]–[15] . To reduce computational overhead, the authors of [9] proposed a JSCC meth

FAQs

What do the experiments indicate about where to apply DSConv replacements?

Replacing Conv with DSConv at intermediate layers of the encoder and decoder yields a favorable complexity-performance trade-off, suggesting layer-wise redundancy in DL-based JSCC.

Jul 16, 2026

Upload by Ava Thompson