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

Area of Science:

• Optics and Photonics
• Computer Vision
• Robotics and Autonomous Systems

Background:

• Underwater optical camera communication (UOCC) performance is limited by fixed camera parameters struggling with dynamic aquatic conditions.
• Image quality, specifically stripe clarity, directly impacts the optical signal-to-noise ratio (SNR) in UOCC systems.
• Varying turbidity, flow velocity, and ambient light significantly degrade UOCC signal reception.

Purpose of the Study:

• To develop a multimodal deep model for predicting scene-optimal camera parameters at capture time for UOCC.
• To enhance imaging clarity, stability, and optical SNR in challenging underwater environments.
• To provide a computationally efficient and portable solution for real-time UOCC.

Main Methods:

• A ResNet50 backbone was used to extract semantic features from raw images.
• Environmental factors (turbidity, flow speed, illumination, LED power) were encoded via a parallel embedding architecture.
• A regression network fused visual and environmental data to predict optimal camera settings (exposure, ISO).

Main Results:

• The proposed model demonstrated robust accuracy in predicting optimal camera parameters.
• Capture-time parameter optimization significantly improved stripe visibility and achieved an average ~3 dB gain in optical SNR.
• The method outperformed fixed camera settings and a post-processing baseline (DnCNN) in SNR stability and performance.

Conclusions:

• The multimodal deep model offers a practical solution for real-time, resource-constrained UOCC.
• Context-aware, capture-time parameter prediction enhances UOCC robustness and image quality.
• This approach enables higher and more stable optical SNR in diverse underwater conditions.