MSR2N: Multi-Stage Rotational Region Based Network for Arbitrary-Oriented Ship Detection in SAR Images

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Synthetic Aperture Radar (SAR) ship detection is challenging due to arbitrary orientations and dense arrangements.
• Existing horizontal bounding box methods lead to redundant detections and missed targets in crowded scenes.
• High Intersection-over-Union (IoU) values between densely packed ships complicate detection with traditional approaches.

Purpose of the Study:

• To develop a robust and accurate ship detection method for SAR images.
• To address limitations of horizontal bounding boxes in detecting arbitrarily oriented and densely packed ships.
• To improve detection performance by reducing background noise and false positives.

Main Methods:

• A Multi-Stage Rotational Region based Network (MSR2N) is proposed, employing rotated bounding boxes.
• The network integrates a Feature Pyramid Network (FPN) for multi-scale feature fusion.
• A Rotational Region Proposal Network (RRPN) with rotation-angle-dependent anchors generates suitable proposals.
• A Multi-Stage Rotational Detection Network (MSRDN) refines proposals through staged regression with increasing IoU thresholds.

Main Results:

• MSR2N effectively utilizes rotated bounding boxes to better represent arbitrary ship orientations.
• The RRPN with multi-angle anchors improves the felicitousness of ship region representation.
• The MSRDN progressively eliminates false positives using incrementally increasing IoU thresholds.
• Experimental results on the SAR Ship Detection Dataset (SSDD) demonstrate state-of-the-art performance.

Conclusions:

• The proposed MSR2N offers a more suitable and robust solution for ship detection in SAR imagery.
• Rotated bounding boxes significantly mitigate issues related to redundancy and missed detections.
• The staged regression approach in MSRDN effectively refines detection accuracy.