CenterNet-Saccade: Enhancing Sonar Object Detection with Lightweight Global Feature Extraction.

Area of Science:

• Marine technology
• Underwater acoustics
• Artificial intelligence

Background:

• Sonar imaging is crucial for marine and underwater monitoring due to sound wave transmission properties.
• Underwater object detection in sonar images can leverage target-shadow relationships.
• Existing deep learning models face challenges in real-time sonar image analysis.

Purpose of the Study:

• To develop an accurate and real-time sonar object detection network using shadow information.
• To enhance detection capabilities by integrating shadow cues into the deep learning model.
• To improve the efficiency and performance of sonar image analysis for marine applications.

Main Methods:

• A novel lightweight network architecture incorporating an attention mechanism with a global receptive field.
• Development of a ShuffleBlock model adapted to an Hourglass backbone for network efficiency.
• Application of CNN dimension reduction to Multi-Head Self-Attention (MHSA) for improved feature processing.
• Modification of the CenterNet sample distribution strategy for better training.

Main Results:

• The proposed network effectively utilizes shadow information for improved target detection in sonar images.
• The lightweight design significantly reduces computational time, enabling real-time monitoring.
• Experimental results demonstrate superior performance compared to conventional deep learning models.
• The model shows strong potential for practical implementation in ocean monitoring.

Conclusions:

• The developed shadow-information-aided detection network provides a significant advancement in real-time sonar object detection.
• The network's efficiency and accuracy make it highly suitable for marine and underwater monitoring tasks.
• This approach offers a promising solution for enhancing the capabilities of autonomous underwater systems.