A Multiscale Feature Pyramid SAR Ship Detection Network With Robust Background Interference

Synthetic aperture radar (SAR) ship detection is widely used in cutting-edge applications such as environmental protection, traffic monitoring, search, and rescue. Lightweight detection algorithms are more important for practical applications. Although there has been extensive research in this field, there are some problems with the existing lightweight algorithms. For example, it is easy to misjudge targets that are mixed with the background, and the detection effect is not ideal for targets with few samples in the dataset. The root cause of these problems lies in the fact that the useless information in the background is relatively close to the target, and existing algorithms are too simplistic in fusing features at different levels, resulting in algorithms not being robust enough when facing these problems. Therefore, this article proposes a multiscale feature pyramid network (FPN)-based detection network (MFPNet), which introduces a spatial information-focusing module in the feature fusion channel to enhance the target's features to suppress interference information in the background and reduce misjudgment. Then, optimize the FPN and extract the importance of different resolution features based on network contribution to identifying multiscale targets. Experiments have shown that the MFPNet has better detection performance compared to existing algorithms on public datasets.