Redefining Accuracy: Underwater Depth Estimation for Irregular Illumination Scenes.

Area of Science:

• Computer Vision
• Robotics
• Oceanography

Background:

• Underwater depth maps are crucial for 3D visualization, navigation, and exploration.
• Existing depth estimation methods struggle with irregular underwater illumination (low-light and overexposure).
• Poor illumination degrades image contrast and object textures, hindering accurate depth perception.

Purpose of the Study:

• To develop an improved self-supervised monocular depth estimation network for underwater environments.
• To address challenges posed by low-light and overexposed underwater imagery.
• To enhance the accuracy and robustness of underwater 3D spatial information acquisition.

Main Methods:

• Proposed an underwater self-supervised monocular depth estimation network.
• Integrated a Monte Carlo Image Enhancement Module (MC-IEM) for probabilistic enhancement of low-light images.
• Utilized transfer learning to incorporate prior knowledge from large-scale models to refine depth estimation and address overexposure.

Main Results:

• The proposed network demonstrated superior performance on two public underwater datasets.
• MC-IEM effectively tackled uncertainty in low-light conditions, improving object recognition and distance estimation.
• Transfer learning refined loss functions and the depth network, mitigating overexposure issues.

Conclusions:

• The novel network significantly advances underwater depth estimation accuracy.
• The integration of image enhancement and auxiliary depth information provides a robust solution for challenging underwater conditions.
• This approach offers more reliable 3D spatial information for various underwater applications.