Subpixel Feature Pyramid Network for Multiscale Ship Detection in Synthetic Aperture Radar Remote Sensing Images

Synthetic aperture radar (SAR) has been widely used in maritime domain awareness, especially in ship detection, due to the capability of working all-day and all-weather. In the detection of SAR ships, there are significant challenges in sea clutter, complex scenes, and especially for multiscale ships with varying sizes. In contrast to large-scale ships, small-scale ships in SAR images only occupy a few pixels and experience more interference. This leads to current ship detection methods being less effective in detecting multiscale ships. Therefore, a novel multiscale ship detection method based on a subpixel feature pyramid network (SFPN) in SAR ship images is proposed. There are two modules in SFPN: the subpixel fusion module (SFM) and the subpixel textural enhancement module (STEM). In SFM, the high-level feature map is merged with the low-level feature map via subpixel convolution for retaining more abundant channel information and taking advantage of multilevel features. Then, the convolutional block attention module is utilized to enhance the extracted salient features to reduce cluttered channel information after fusion. By these means, the information retention of small targets is better. In STEM, a semantic enhancement module and a textural enhancement module are proposed to provide contextual information for accurately localizing objects and understanding scenes. Finally, the experimental results demonstrate the excellent ship detection performance of SFPN compared to seven feature pyramid network-based (FPN-based) and feature pyramid network-free (FPN-free) state-of-the-art methods. Specifically, the $\text{AP}_{50}$ and $\text{AR}_{50}$ increase of SFPN is 1.7%, 0.2% on SAR-Ship-Dataset, 3.2%, 1.2% on rotated ship detection dataset in SAR images (RSDD-SAR).