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

Ships in optical remote sensing imagery are often affected by significant background noise, and they typically appear as small targets with large aspect ratios, dense distributions, and arbitrary orientations, which pose substantial challenges for accurate detection. To address these issues, this paper proposes a novel detection network, termed TCA-Net, specifically designed for ship detection in complex remote sensing scenarios. TCA-Net comprises four main modules: a Deformable Attention Pyramid Network (DAPNet), a Multi-Scale Feature Enhancement Network (MFENet), a Multi-Scale Adaptive Pooling Network (MAPNet), and a Rotation Head. First, DAPNet integrates a custom-designed Ship Deformable Convolution unit and Ship Attention modules to effectively capture features of ships with arbitrary shapes and orientations while suppressing multi-scale background noise. Second, MFENet employs parallel Laplacian and dilated convolutions to enhance features of ships across various scales, particularly improving the detection of small targets. Third, MAPNet adaptively generates precise ROI for ships with large aspect ratios and learns multi-scale contextual information to infer relationships in crowded scenes. Finally, the Rotation Head regresses rotated bounding boxes to precisely localize ships at any orientation. Experiments conducted on the public DOTA and HRSC2016 datasets demonstrate that TCA-Net achieves SOTA performance.