A Low-Power Semantic Communication Framework for Multimodal Fusion via Non-Binary Polar Codes

Multimodal semantic communication enables efficient and accurate cross-modal information transmission, yet precisely integrating semantic information from different modalities in dynamic communication environments remains challenging. To address this, this paper proposes a non-binary polar coding and decoding scheme for multimodal semantic feature fusion. This approach limits channel coding during the training phase to reduce overhead, while leveraging the high coding gain of non-binary polar codes and a deep neural network-based fast decoder to decrease complexity and power consumption. In the semantic module, we design a feature fusion network based on attention mechanisms and residual modules to more accurately capture cross-modal semantic features. For transmission, an adaptive non-binary polar (ANP) encoding strategy is introduced, while a neural network-driven non-binary polar (DNP) decoder is employed in the decoding part. Simulation results demonstrate that the proposed system significantly improves transmission efficiency and reliability while reducing bit error rate, and effectively lowers transmission power consumption.