A robust network for tiny and arbitrary-oriented ship detection in remote sensing images.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Ship detection in optical remote sensing is challenging due to background noise, small target size, dense distribution, and arbitrary orientations.
• Existing methods struggle with these complex scenarios, necessitating advanced detection techniques.

Purpose of the Study:

• To propose a novel detection network, TCA-Net, specifically designed for accurate ship detection in challenging remote sensing imagery.
• To address issues of background noise, arbitrary orientations, and scale variations in ship detection.

Main Methods:

• TCA-Net integrates four modules: Deformable Attention Pyramid Network (DAPNet), Multi-Scale Feature Enhancement Network (MFENet), Multi-Scale Adaptive Pooling Network (MAPNet), and a Rotation Head.
• DAPNet uses deformable convolutions and attention to capture ship features and suppress noise.
• MFENet enhances multi-scale features for small targets, while MAPNet handles large aspect ratios and crowded scenes.
• The Rotation Head enables precise localization of ships with rotated bounding boxes.

Main Results:

• Experiments on DOTA and HRSC2016 datasets show TCA-Net achieves state-of-the-art (SOTA) performance.
• The proposed network effectively detects ships with arbitrary orientations and suppresses background noise.
• Improved detection of small targets and ships in crowded scenes was observed.

Conclusions:

• TCA-Net demonstrates superior performance in ship detection within complex remote sensing scenarios.
• The network's modular design effectively addresses key challenges, offering a robust solution for maritime surveillance and analysis.
• The findings highlight the potential of advanced deep learning architectures for remote sensing applications.