A multi-scale rotated ship targets detection network for remote sensing images in complex scenarios

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Detecting small, rotated ship targets in complex remote sensing images is challenging due to limited features and orientation variations.
• Existing object detection algorithms often struggle with missed and false detections in such scenarios.

Purpose of the Study:

• To propose a Multi-Scale Rotated Detection Network (MSRO-Net) for accurate detection of rotated ship targets in remote sensing imagery.
• To enhance the network's ability to extract features from targets with limited information and random orientations.

Main Methods:

• Developed a CNN-Transformer hybrid architecture for collaborative feature extraction, integrating a Coordinate-Aware Pyramid Feature Aggregation (CAPP) module.
• Introduced an Upsampling Feature Reconstruction Pyramid (ARFPN-C) utilizing Adaptive Rotated Convolution (ARC) for improved perceptual capabilities and multi-scale feature fusion.

Main Results:

• MSRO-Net achieved superior performance on HRSC2016 and DOTA datasets.
• Demonstrated high accuracy with mAP07 of 90.70%, mAP12 of 98.98% on HRSC2016, and 89.46% mAP for ship detection in DOTA.

Conclusions:

• The proposed MSRO-Net effectively addresses challenges in detecting rotated ship targets in remote sensing images.
• The network's hybrid architecture and novel modules significantly improve detection accuracy and robustness.