Underwater Rescue Target Detection Based on Acoustic Images

Area of Science:

• Robotics and Autonomous Systems
• Artificial Intelligence
• Signal Processing

Background:

• Effective underwater search and rescue is critical for drowning victims in water emergencies.
• Traditional methods are limited by complex underwater environments and low visibility.
• Unmanned Underwater Vehicles (UUVs) with sonar and deep learning offer improved efficiency for active searches.

Purpose of the Study:

• To develop an advanced algorithm for underwater acoustic rescue target detection, specifically for small targets.
• To address the challenge of cross-domain adaptability in deep transfer learning for acoustic images.
• To create a lightweight, efficient model suitable for embedded systems on underwater UAVs.

Main Methods:

• Construction of a sound-based rescue target dataset using deep transfer learning.
• Proposal of a two-branch convolution module and improved YOLOv5s algorithm for small acoustic target detection.
• Development of a heterogeneous information hierarchical migration learning method, freezing network layers to enhance accuracy.
• Integration of ShuffleNetv2 for a lightweight model applicable to underwater UAVs.

Main Results:

• The proposed heterogeneous information hierarchical migration learning method demonstrated superior cross-domain adaptability compared to direct fine-tuning.
• The improved YOLOv5s model with the two-branch convolution module significantly enhanced detection accuracy for small acoustic targets.
• The lightweight ShuffleNetv2-based model achieved state-of-the-art performance in underwater search and rescue target detection tasks.
• Extensive comparative experiments validated the feasibility and effectiveness of the developed methods on various acoustic images.

Conclusions:

• The developed deep learning approach offers a significant advancement in underwater acoustic target detection for rescue operations.
• The heterogeneous information hierarchical migration learning method effectively overcomes domain shift issues between optical and acoustic data.
• The lightweight model is well-suited for real-time deployment on resource-constrained underwater UAVs, enhancing search and rescue capabilities.