SAR Ship Detection Method Based on Convolutional Neural Network and Multi-layer Feature Fusion

SAR ship detection plays an important role in marine traffic monitoring. Traditional SAR target detection methods are mostly based on intensity differences between target and clutter, which is limited especially in complex scenes, for instance coastal areas. In order to improve the detection performance in complex scenes, a SAR ship detection method based on convolutional neural network named LCMF is proposed in this paper. Firstly, a base network with low complexity is employed to extract features. Secondly, the ‘top-down’ approach is adopted to gradually fuse the semantically strong features, which is helpful for reducing false alarms, with the low-level high-resolution features to improve the detection performance on small targets. Finally, small-scaled anchor is designed to obtain region proposals, and these proposals are further fed to classification and regression network, which outputs the final detection results. Experiments on the sentinel-1A dataset demonstrate that the proposed method can detect ship targets in SAR images of complex scenes with high speed and accuracy.