RT-DETR-FCES: a lightweight ship detection algorithm from remote sensing perspective.

Area of Science:

• Computer Vision
• Remote Sensing
• Maritime Surveillance

Background:

• Ship detection is vital for maritime safety and management.
• Existing methods struggle with accuracy, real-time performance, and lightweight deployment, especially for small targets in complex environments.
• Resource-constrained platforms pose additional challenges for effective ship detection.

Purpose of the Study:

• To develop a lightweight, real-time ship detection framework for optical remote sensing images.
• To address the limitations of current methods in balancing accuracy, speed, and model size.
• To improve the detection of small ship targets on platforms with limited computational resources.

Main Methods:

• Proposed RT-DETR-FCES framework utilizing FasterNet as the backbone for reduced computational cost.
• Incorporated Cascaded Group Attention (CGA) for enhanced multi-scale feature fusion.
• Introduced an EUCB-SC module to boost feature representation and inference efficiency.

Main Results:

• Achieved 87.5% mAP on the MASATI-v2 dataset and 49.1% mAP on the VAI dataset.
• Reduced model parameters by 45.5% and FLOPs by 48.0% compared to the baseline RT-DETR.
• Attained high inference speeds of 87.7 FPS and 85.1 FPS, demonstrating real-time capability.

Conclusions:

• The RT-DETR-FCES framework effectively balances detection accuracy and lightweight efficiency.
• Demonstrated significant improvements in computational cost and inference speed.
• Proven effective for real-time maritime ship monitoring applications, especially on resource-constrained systems.