Image Semantic Segmentation of Underwater Garbage with Modified U-Net Architecture Model

Area of Science:

• Robotics
• Computer Vision
• Environmental Science

Background:

• Autonomous underwater garbage collection faces significant challenges in object localization and recognition.
• Effective identification of debris is crucial for robotic environmental remediation missions.

Purpose of the Study:

• To develop an efficient deep learning model for semantic segmentation of underwater garbage.
• To create a specialized dataset for training and evaluating underwater garbage recognition systems.
• To optimize model performance by analyzing various hyperparameters, loss functions, and optimizers.

Main Methods:

• A modified U-Net architecture with a deeper contracting path and expansive path was designed for end-to-end image semantic segmentation.
• A novel dataset specifically for underwater garbage semantic segmentation was established.
• Extensive experiments were conducted to evaluate the model's performance, including the impact of focal loss on class imbalance.

Main Results:

• The modified U-Net architecture demonstrated effective performance in segmenting underwater garbage.
• The focal loss function significantly improved results by addressing the target-background imbalance problem.
• The established dataset facilitated robust model verification and hyperparameter tuning.

Conclusions:

• The proposed modified U-Net architecture provides a robust solution for underwater garbage semantic segmentation.
• The findings offer a foundational contribution to enhancing the capabilities of underwater robots for environmental cleanup.
• Precise underwater target recognition is achievable, paving the way for more effective robotic operations.