YOLOv5s-CA: A Modified YOLOv5s Network with Coordinate Attention for Underwater Target Detection

Area of Science:

• Computer Vision
• Robotics
• Marine Technology

Background:

• Underwater target detection is crucial for marine applications like surveillance and rescue.
• Current algorithms struggle with complex environments and limited data, leading to unsatisfactory accuracy.
• Improving underwater target recognition is essential for advancing marine robotics and exploration.

Purpose of the Study:

• To enhance the accuracy of underwater target detection algorithms.
• To propose a modified YOLOv5s network incorporating attention mechanisms for improved performance.
• To address the limitations of existing methods in complex underwater conditions.

Main Methods:

• A modified YOLOv5s network, termed YOLOv5s-CA, was developed by integrating Coordinate Attention (CA) and Squeeze-and-Excitation (SE) modules.
• The number of bottlenecks in the initial C3 module was increased to enhance shallow feature extraction.
• CA modules were embedded within C3 modules, and SE layers were added to the output for focused attention.

Main Results:

• The YOLOv5s-CA network demonstrated improved feature extraction and attention capabilities.
• Experiments conducted on the 2019 China Underwater Robot Competition dataset showed significant performance gains.
• The modified network achieved a 2.4% increase in mean Average Precision (mAP) compared to the baseline YOLOv5s.

Conclusions:

• The proposed YOLOv5s-CA network effectively improves underwater target detection accuracy.
• Attention mechanisms and enhanced feature extraction are key to overcoming challenges in underwater environments.
• This research contributes to more reliable underwater target recognition for various marine applications.