Lightweight Semantic Segmentation Network for End-to-End Learning-based Communication

This is the authors' abstract. We don't add key points for this paper.

Semantic communication has received growing interest since it can remarkably reduce the amount of data to be transmitted without missing critical information. However, tackling bandwidth limitations often incurs excessive computational overhead and slows the inference speed. Instead of relying on compression, coding, or block-wise optimization, we propose the Multi-Modal Channel Fusion Network (MMCFNet), a lightweight end-to-end semantic segmentation network for multi-modal channel fusion. MMCFNet enhances image segmentation performance, effectively reduces energy consumption, and demonstrates strong noise resilience. We proposed a Deep Joint Source-Channel Coding (DJSCC) channel fusion framework with multi-modal information fusion capabilities and superior noise immunity to further improve efficiency. Additionally, MMCFNet effectively supports multi-modal information processing while maintaining robustness to noise and interference. Validated on various benchmark datasets, MMCFNet achieves low computational overhead and significantly outperforms state-of-the-art methods on MFNet, PST900, and MCubeS. The proposed framework demonstrates robustness to noise and interference, making it ideal for bandwidth-constrained semantic communication systems.