Advanced deep learning framework for underwater object detection with multibeam forward-looking sonar.

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Marine Engineering

Background:

• Underwater object detection (UOD) is crucial for infrastructure maintenance and asset management.
• Sonar imaging is preferred over optical methods in challenging underwater conditions.
• Existing sonar-based UOD methods struggle with low resolution and poor contrast, limiting precision and transferability.

Purpose of the Study:

• To develop an advanced deep learning framework for improved UOD using multibeam forward-looking sonar data.
• To address the precision and transferability challenges in current sonar-based object detection algorithms.

Main Methods:

• Adapted the YOLOv7 network architecture for sonar data.
• Implemented unique optimizations in data preprocessing, feature fusion, and loss functions.
• Validated the framework on a public dataset and through experiments with an underwater remotely operated vehicle.

Main Results:

• The proposed framework demonstrated superior object classification performance compared to existing sonar-based methods.
• Achieved significant enhancements in target classification, localization, and transfer learning capabilities.
• Experimental validation confirmed the framework's effectiveness on an underwater remotely operated vehicle.

Conclusions:

• The advanced deep learning framework offers a significant improvement for UOD using sonar data.
• The framework shows potential for underwater structural inspection, monitoring, and autonomous asset management due to its enhanced capabilities.
• This work advances the application of deep learning in marine robotics and infrastructure management.