UGC-Net: a generative communication framework for underwater images oriented to semantic understanding

In marine exploration, underwater images represent the most intuitive means for human perception of the underwater environment. However, traditional image encoding and transmission methods struggle to ensure clear reconstruction quality due to bandwidth limitations and significant interference in underwater acoustic (UWA) channels. Semantic communication seeks to transform pixel-level transmission into semantic-level transmission, thereby enhancing the reliability and efficiency of communication. In this paper, we propose a generative underwater image communication framework designed for semantic understanding, referred to as UGC-Net. In this framework, only semantic features are reliably transmitted end-to-end through the UWA channels. At the receiver, the spatial relationships between semantic features are first learned based on contextual information, which is then used to generate high-quality underwater images employing a conditional diffusion model. Experimental results demonstrate that our approach exhibits robust and accurate generative performance, even in challenging conditions characterized by low signal-to-noise ratios (SNR).