Bidimensional Joint Mask Learning Enabled Multi-Rate Semantic Communications
Mingtong Zhang, Haixia Zhang, Wenjie Liu, Hui Ding, Dongfeng Yuan · IEEE Wireless Communications Letters · 2025
Deep learning (DL)-based semantic communications have gained significant attention for enhancing communication efficiency. However, existing DL-based schemes typically operate with a fixed semantic coding rate (SCR), limiting flexibility across varying channel environments and diverse user demands. To address this limitation, this paper proposes a novel bidimensional joint mask learning enabled multi-rate semantic communication (BiMSC) scheme. Specifically, an effective bidimensional joint mask mechanism is first designed to achieve fine-grained importance distinction among semantic code symbols. Then, a feature-aware network (FANet) is developed to dynamically determine the optimal mask length at different dimensions by jointly considering SCR, channel conditions, and data contents. Through end-to-end learning, arbitrary SCRs with high semantic recovery performance can be achieved. Simulation results demonstrate that the proposed BiMSC scheme exhibits robustness across different channel conditions, and outperforms other multi-rate semantic communication schemes in terms of BLEU score.