BD-RDE: Bridging Domains for Robust Depth Estimation in Underwater Environments with a Color-Balance Domain

Accurate 3D environment perception is essential for effective underwater robot control and human-robot collaboration, allowing robots to plan safe navigation and interact efficiently with human operators. However, underwater images are often distorted by light scattering and absorption, resulting in color distortion and blurred textures that pose challenges for depth-estimation accuracy. Underwater sensors, such as radar and stereo cameras, suffer from limited range and sensitivity owing to light attenuation in marine environments. To improve the depth-estimation accuracy in underwater environments, this study introduces a novel self-supervised enhancement module inspired by domain-adaptation techniques. The module includes a color balance domain to bridge both the normal and underwater domains, along with an edge-aware module to better capture texture details. This enabled model training without requiring additional underwater images. Our enhancement module seamlessly integrates with existing frameworks, significantly enhancing performance under challenging underwater conditions. Compared with the baseline model, the enhanced depth estimation framework demonstrated a substantial average residual mean square error decrease of 20.4% across ten different levels of degraded underwater images. Our experimental results validate the effectiveness of the approach for underwater depth estimation and its superior generalization, thus opening new possibilities for improved underwater robot navigation and human-computer interaction.