Domain Knowledge Driven Semantic Communication for Image Transmission Over Wireless Channels

Recent progress in semantic communication has shown potential in improving transmission efficiency by the removal of signal-level redundancies. However, existing semantic communication systems fail to consider the high-level conceptual semantics embedded in the source, which is critical for human perception especially in certain domains. This letter proposes a novel domain knowledge driven semantic communication (DKSC) system, which is a dual-path framework with semantic extraction and reconstruction at both information and concept levels. Specifically, the dual-path semantic encoder and decoder are jointly designed and optimized in the end-to-end wireless communication system based on deep neural networks. To illustrate the performance of DKSC system, we execute image transmission task over wireless channels in medical domain as an example, and the impacts of various channel models and compression ratios are investigated. Experimental results demonstrate that with the integration of domain knowledge semantics, the proposed DKSC system outperforms other baseline schemes in terms of image restoration metrics with robustness against channel variations.