Cross-Modal Semantic Fusion and Compression for Multimodal Semantic Communication
Year: 2026
Type: conference-paper
Abstract: Semantic communication has emerged as a paradigm shift in information transmission, enabled by semanticlevel compression. However, existing multimodal approaches struggle with heterogeneous data transmission and fail to fully exploit inter-modal semantic correlations, resulting in limited compression and reconstruction accuracy. To this end, this paper proposes a multimodal semantic communication method with Transformer-driven fusion compression. We design a multimodal compression semantic neural network framework for semantic communication, and adopt a cross-attention mechanism to capture semantic relationships between text and image modalities, and introduce an image-oriented masked autoencoder with an asymmetric encoder-decoder structure to suppress image semantic dominance and reduce redundant information during fusion. Experimental results on multiple multimodal datasets demonstrate that our proposed method can achieve superior data reconstruction accuracy and compression efficiency, as well as improved model robustness under varying signal-to-noise ratio conditions, compared with existing models.
Institutions: Northwest University, Northwest University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7220938643