Multi-Level Similarity for Efficient Compression in Graph-Based Multi-Modal Semantic Communication

Efficient and accurate multi-modal data transmission remains a key challenge in semantic communications. Existing research has overlooked the multi-level semantic similarity inherent in multi-modal data. Motivated by this, a graph-based multi-modal semantic communication system (GMSC) is proposed. In GMSC, triplets fully correlate intra-modal and cross-modal semantic similarities, enabling the integrated extraction of redundant semantic information. A task-oriented compression threshold method is designed to remove redundant transmission information and enhance efficiency. Simulation results demonstrate that GMSC achieves a 7% improvement in task accuracy at the same compression threshold and significantly improves transmission efficiency under the same task accuracy.