Topology-aware Interpretable Image Semantic Communication

In the 6 G era, semantic communication (SemCom), which transmits high-level semantics instead of bit-wise data, shows great potential to improve transmission efficiency and robustness. However, existing methods often rely on deep feature extraction, lacking interpretability and may lead to redundant transmission. This paper proposes an interpretable SemCom framework based on scene graphs. Firstly, we use explicit scene graphs to represent semantic content, and a topology-aware codec to efficiently compress semantic structures while preserving topological relationships. To verify semantic consistency, the reconstructed scene graph is further utilized to generate images in various styles using a large language model (LLM). Simulation results show that, compared with traditional SemCom approaches, the proposed method achieves a $\mathbf{1 2. 6 \%}$ improvement in semantic consistency score (SCS), a $\mathbf{1 0. 2 \%}$ increase in structural similarity index (SSIM), and a $\mathbf{3 2. 5 \%}$ reduction in graph edit distance (GED), with a compression rate exceeding 95%. These results highlight the ability of the framework to balance efficiency and semantic fidelity, which offers an effective and interpretable solution for future SemCom.