Flexible Joint Source-Channel Coding for Image Semantic Communication
Wenyuan Wang, Chaowei Wang, Jisong Xu, Lexi Xu, Mingliang Pang, Changxin Lu, Ziyi Huang, Weidong Wang, Peng Yu, Fan Jiang, Yinghui Ye · 2024
Semantic communication aims to ensure the successful delivery of information meaning and is expected to become one of the potential technologies for next-generation communications. However, existing approaches to semantic communication systems often draw inspiration from computer vision, focusing predominantly on transmission accuracy while less attention is given to overall efficiency. In this paper, we propose a flexible joint source-channel coding scheme (FJSCC) that comprehensively considers reconstruction accuracy and latency. The study emphasizes the dynamic adjustment of neural network depth based on channel state information (CSI) to optimize both the coding efficiency and the semantic information extraction in variable channel conditions. Additionally, we propose a simple scheme for compressing semantic features, utilizing the dispersed distribution of semantic features in feature maps to train the neural network. The proposed scheme shows performance improvements over existing algorithms. By using semantic communication utility metric, we compare the efficiency differences of FJSCC at various depths and compression ratios (CRs), laying the groundwork for future research on the efficiency of semantic communications.