【摘要】 针对6G“AI(Artificial Intelligence)与通信融合”场景下，经典香农信息论以像素级误差为优化准则难以满足人类主观感知需求，且现有感知图像传输缺乏统一理论指导的问题。从语义信息论的同义性视角出发，提出一种面向语义通信的同义变分推断理论。该方法将感知相似性作为典型的同义准则来物理建构理想同义集，并在语法层面通过最小化部分语义KL(Kullback-Leibler)散度来近似隐同义表征的真实后验分布。首先，在严格的数学框架下推导了同义变分推断的等效似然优化方向，从理论上证明了感知图像语义传输遵循一种“同义率-失真-感知”的三元权衡界限；随后，基于该理论界限设计了一种渐进式同义图像编码框架，通过构建多层级同义表征，发送端可以根据编码或通信的场景需求，仅对特定同义层级下的同义表征进行编码和传输，而接收端则可以在重构同义集内通过采样实现细节补全。实验结果表明，所提理论推导涵盖了现有的率失真感知经验方案，且通过单一的渐进式SIC(Synonymous Image Compression)模型在多种编码速率下，实现渐进式地主观感知质量恢复，从而为6G智能媒体传输提供了有效的基础理论与技术路径。更多还原

【Abstract】 In 6G scenarios characterized by the integration of AI and communications, classical Shannon information theory, which uses pixel-level error as the optimization criterion, can hardly satisfy human subjective perceptual requirements. Moreover, existing perceptual image transmission schemes lack unified theoretical guidance. From the perspective of synonymy in semantic information theory, this paper proposes a synonymous variational inference theory for semantic communications. The proposed method physically constructs an ideal synonymous set by taking perceptual similarity as a typical synonymous criterion, and approximates the true posterior distribution of latent synonymous representations at the syntactic level by minimizing the partial semantic Kullback-Leibler divergence. First, under a rigorous mathematical framework, the equivalent likelihood optimization direction of synonymous variational inference is derived, theoretically proving that perceptual image semantic transmission follows a ternary tradeoff bound among synonymous rate, distortion, and perception. Then, based on this theoretical bound, a progressive synonymous image coding framework is designed. By constructing multi-level synonymous representations, the transmitter can encode and transmit only the synonymous representation at a specific synonymous level according to the coding or communication scenario requirements, while the receiver can complete details within the reconstructed synonymous set through sampling. Experimental results show that the proposed theoretical derivation covers existing empirical rate-distortion-perception schemes. Moreover, with a single progressive synonymous image compression model, progressive recovery of subjective perceptual quality is achieved under multiple coding rates, thereby providing an effective fundamental theory and technical path for intelligent media transmission in 6G.更多还原