MIMO Semantic Communication via Score-Based Reverse Mean Propagation

This paper introduces a novel multiple-input multiple-output (MIMO) semantic communication system by leveraging advanced score-based generative model. The proposed system features a deep neural network (DNN)-based encoder at the transmitter and a score-based decoder at the receiver, where the latter directly reconstructs source data from channel outputs without intermediate signal detection. We establish a unified Bayesian framework for tranceiver design rooted in the information-maximization principle. Based on this framework, we formulate a variational inference problem for receiver design and develop a principled score-based decoding algorithm, which generalizes reverse mean propagation (RMP) by incorporating the coding and channel constraints in likelihood calculations. Experimental results on the FFHQ dataset demonstrate significant improvements in perceptual metrics over existing benchmarks, validating the effectiveness of the proposed approach in achieving high-fidelity image transmission.