Implicit codebook-based noise resilient underwater digital semantic communication system

State-of-the-art underwater optical semantic communication (UWOSC) systems achieve digital compatibility through vector quantization (VQ) with explicit codebooks. While efficient, this method suffers from structural mismatches between discrete codebook indices and continuous feature representations, compounded by the codebook collapse issues. In this paper, we introduce what we believe to be a novel approach to address these challenges by developing a robust implicit codebook-based UWOSC system for image transmission. Specifically, we utilize a CNN-based semantic feature transceiver to extract image semantic features, and employ a finite scalar quantization (FSQ)-based implicit codebook construction method with dimension-wise bit mapping transmission design. Moreover, adversarial training is adopted to refine the visual fidelity of reconstructed images through a patch discriminator. The robustness primarily stems from its dimension-wise bit mapping design. Furthermore, the predefined structure of our codebook space stabilizes network training and prevents codebook collapse. To further enhance the robustness of the model, we propose a dual-phase training strategy. Experiment results demonstrate that the FSQ-based implicit codebook-based approach outperforms VQ-based explicit codebook methods, delivering superior performance with reduced communication overhead, particularly under challenging underwater wireless optical communication (UWOC) channel conditions.