SemFusion: Multi-source semantic information fusion and communication

This paper puts forth SemFusion, a semantic information fusion and communication framework for two-hop multisource relay networks. Although deep learning-based semantic communication is considered a new paradigm for next-generation communication networks, most prior works have focused on single-source scenarios, especially in relay networks. In contrast, we investigate a multi-source scenario where multiple sensors monitor the same scene from different angles and send partial images to the destination via a relay. The destination receives partial images of the monitored scene to reconstruct the complete image. Empowered by semantic communication, in the first hop, SemFusion allows only a subset of sensors to send their semantic information of the partial images to the relay. In the second hop, instead of forwarding the semantic information of each sensor separately, the relay further performs semantic information fusion so that only the most valuable semantic information is sent to the destination. Moreover, in contrast to the conventional end-to-end training method used in semantic communication, we propose a two-stage training strategy, where each stage corresponds to one hop, to improve the training efficiency of SemFusion. Experiments indicate that SemFusion significantly saves communication resources and provides better image reconstruction quality than state-of-the-art semantic forwarding strategies. Copyright © 2024 IEEE.