YOLO-CPCL: Compact Multi-Class Oriented Ship Detection with Adaptive Feature Fusion and Aspect-Ratio-Coupled Angle...

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Ship detection in optical remote sensing images is difficult due to dense arrangements, nearshore environments, and complex ship appearances.
• Elongated hulls, varied orientations, and background clutter challenge feature aggregation and accurate rotated-box localization.

Purpose of the Study:

• To develop an improved oriented ship detection model for complex remote sensing scenes.
• To enhance feature aggregation and angle supervision for better detection of elongated and arbitrarily oriented ships.
• To reduce model complexity and computational cost without sacrificing performance.

Main Methods:

• Proposed YOLO-CPCL, a compact oriented detector based on YOLOv8n-OBB.
• Introduced a Ship-Oriented Slender Adaptive Fusion module (SOSA-Fuse) to enhance feature aggregation for slender targets.
• Developed an Aspect-Ratio-Coupled Angle Supervision Loss (ARCAS-Loss) for improved angle supervision.

Main Results:

• YOLO-CPCL achieved significant improvements in precision, recall, and mean Average Precision (mAP) on the HRSC2016 dataset.
• Reduced parameter count by 5.19% and computational cost (GFLOPs) by 8.43%.
• Demonstrated positive performance gains on ShipRSImageNet and DOTA-v1.0 datasets.

Conclusions:

• The proposed YOLO-CPCL method effectively improves ship recall and high-IoU oriented localization in challenging remote sensing scenarios.
• The model achieves superior performance with reduced computational resources, making it suitable for practical applications.
• SOSA-Fuse and ARCAS-Loss modules contribute to enhanced feature aggregation and robust angle supervision for oriented ship detection.