CCW-YOLOv5: A forward-looking sonar target method based on coordinate convolution and modified boundary frame loss

Area of Science:

• Marine technology
• Artificial intelligence
• Computer vision

Background:

• Multi beam forward looking sonar (MFLS) is crucial for underwater detection.
• Complex environments and noise limit MFLS recognition performance.
• Existing MFLS systems struggle with unclear features and scarce data.

Purpose of the Study:

• To improve YOLOv5 for enhanced underwater target detection using MFLS data.
• To address challenges of limited sonar data and feature extraction.
• To increase the accuracy and generalization of MFLS target recognition.

Main Methods:

• Adapted YOLOv5 architecture with transfer learning to handle limited sonar image data.
• Integrated coordinate convolution to enhance extraction of positional features for small targets.
• Incorporated attention mechanisms to optimize feature learning and expand receptive fields.
• Modified the bounding box loss function to mitigate issues with uneven training sample quality.

Main Results:

• The proposed CCW-YOLOv5 algorithm achieved 85.3% mAP@0.5 for object detection.
• Demonstrated a fastest inference speed of 54 FPS on local machine.
• Showcased significant improvements in detection accuracy and performance over existing algorithms.

Conclusions:

• The CCW-YOLOv5 algorithm offers superior performance for underwater target detection with MFLS.
• The enhancements improve model applicability and generalization across diverse underwater conditions.
• This research provides a robust solution for challenging MFLS target recognition tasks.