Multimodal deep learning for intelligent camera parameter control in underwater optical camera communication imaging

Underwater optical camera communication (UOCC) relies on image-based signal reception, where the quality of recorded stripes directly determines the achievable optical signal-to-noise ratio (SNR). However, fixed camera parameters-such as exposure level and International Standards Organization (ISO) sensitivity-are often inadequate under dynamically varying aquatic conditions including turbidity, flow velocity, and ambient illumination. To overcome these limitations, we propose a multimodal deep model that fuses visual cues with environmental context to predict scene-optimal camera parameters at capture time. A ResNet50 backbone extracts semantic representations from raw stripe images, while environmental factors-including turbidity, flow speed, ambient illumination, and LED power-are jointly encoded through a parallel embedding architecture. These modalities are fused within a regression network to infer settings that maximize imaging clarity and stability. The experimental results demonstrate that the proposed model achieves robust accuracy. Crucially, the capture-time parameter selection improves stripe visibility and delivers an average ∼3dB gain in optical SNR across diverse conditions. This capture-time optimization sustains a higher and more stable SNR band than both the original fixed settings and a representative learning-based post-processing baseline (DnCNN with horizontal-banding suppression). Beyond accuracy, the approach is computationally light and portable: context-aware parameter prediction at capture time eliminates per-frame processing, providing a practical route to real-time, resource-constrained UOCC with enhanced image quality and robustness.