DRC²-Net: A Context-Aware and Geometry-Adaptive Network for Lightweight SAR Ship Detection.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Synthetic Aperture Radar (SAR) ship detection faces challenges from background clutter, target sparsity, and occluded ships, especially at small scales.
• Existing methods struggle with adaptive feature extraction for varying geometric deformations and scale variations in complex SAR imagery.

Purpose of the Study:

• To develop a lightweight and efficient SAR ship detection framework that addresses challenges in complex environments.
• To enhance contextual awareness and improve adaptive feature extraction for multi-scale ship detection.

Main Methods:

• Proposed the Deformable Recurrent Criss-Cross Attention Network (DRC2-Net), a YOLOX-Tiny based framework.
• Incorporated Recurrent Criss-Cross Attention (RCCA) for contextual awareness and Deformable Convolutional Networks v2 (DCNv2) for adaptive feature extraction.
• Trained and evaluated the model on the SSDD and iVision-MRSSD datasets, featuring diverse SAR imagery.

Main Results:

• DRC2-Net achieved superior performance over the YOLOX-Tiny baseline on the SSDD dataset, with notable improvements in AP@50, APs, APm, and APl.
• Demonstrated enhanced scale-aware detection capabilities on the challenging iVision-MRSSD dataset, outperforming state-of-the-art detectors across various target scales.
• The compact model (5.05 M parameters) ensures strong generalization and real-time applicability.

Conclusions:

• DRC2-Net effectively addresses the limitations of current SAR ship detection methods, particularly for small, fragmented, or occluded targets.
• The proposed architecture offers a robust and efficient solution for multi-scale ship detection in complex SAR environments.
• The model's adaptive feature extraction and enhanced contextual awareness contribute to its superior performance and generalization capabilities.