Deep Joint Source-Channel Coding for Wireless Image Transmission with Adaptive Models

The implementation of joint source-channel coding (JSCC) schemes using deep learning has accelerated the development of semantic communication research. Existing JSCC schemes based on deep learning (DL) are trained on a fixed signal-to-noise ratio (SNR); however, these trained models are not designed for scenarios in which the SNR is dynamic. Therefore, a novel semantic adaptive model for semantic communication—called joint source-channel coding with adaptive models (AMJSCC)—that has a semantic adaptive model selection (SAMS) module is proposed. The joint source-channel encoding (JSCE) model and the joint source-channel decoding (JSCD) model adapt according to both real-time channel conditions and system available computational power resources. Furthermore, residual networks with different layers are investigated to further improve the accuracy of information recovery. Simulation results demonstrate that our model can achieve higher recovery similarity and is more robust and adaptive to the SNR and communication resources. Meanwhile, compared to the state-of-the-art deep JSCC methods, it reduces storage space and communication resource consumption