Advancing monocular depth estimation by integrating underwater optical imaging priors into transformer-based network.

Area of Science:

• Computer Vision
• Robotics
• Optical Imaging

Background:

• Underwater monocular depth estimation is crucial for autonomous robots but faces challenges due to image degradation from light scattering and absorption.
• Existing methods struggle with poor contrast and loss of detail in underwater imagery.

Purpose of the Study:

• To enhance the accuracy of underwater monocular depth estimation for resource-constrained miniature autonomous robots.
• To address the challenges posed by degraded underwater image quality.

Main Methods:

• Proposed an underwater light priors processing module (ULPM) integrated with EfficientNet features.
• Utilized a transformer-based network incorporating a physical model of underwater optical imaging.
• Introduced a classification loss alongside regression for robust depth estimation.

Main Results:

• The proposed algorithm achieved state-of-the-art performance on the Flsea and USOD10K datasets.
• Demonstrated superior accuracy and robustness compared to existing underwater depth estimation methods.
• The ULPM and classification loss effectively improved depth estimation in degraded underwater images.

Conclusions:

• The developed method significantly advances underwater monocular depth estimation capabilities.
• The integration of physical underwater imaging models and hybrid loss functions offers a promising direction for future research.
• The approach provides a more accurate and reliable solution for robotic navigation and perception in underwater environments.