Dynamic Convolution and Transformer Based Dual-Branch Coding in Semantic Communication System

Abstract

Semantic communication is a potential key technology in 6G intelligent communication era. To reduce the information redundancy and improve the accuracy of text data transmission, this letter proposes a novel dual-branch joint source-channel coding model in semantic communication system that integrates dynamic convolution (DynConv) with Transformer architecture. The system utilizes the convolutional attention mechanism to capture the local semantic details and the Transformer self-attention mechanism for contextual associations within sentences. In particular, we further investigate the dimension allocation strategy between these two attention mechanisms, seeking the optimal balance between the recovery performance and complexity of the system. Simulation results demonstrate that, compared with the traditional communication system and the typical semantic communication system based on the standard Transformer, the proposed system achieves higher recovery performance, better robustness, and lower complexity.