Ship Detection in Large-scale SAR Images Based on Dense Spatial Attention and Multi-level Feature Fusion

Abstract

In recent years, ship detection in large-scale synthetic aperture radar (SAR) images attracts more attentions and becomes a research hotspot. But it still faces some challenges, such as strong interference of background noise and very small ship targets. This paper proposes a novel anchor-free detector, small target detector (STDet), based on dense spatial attention (DSA) and multi-level feature fusion (MFF). DSA is applied first to the backbone (Resnet50) in order to filter out the background noise and obtain more advanced semantic features. Then, a MFF network is used after the backbone to improve the detection accuracy, especially for small targets, by fusing the location and semantic information of different level feature maps. Finally, the refined features are fed to detection head to get the final results. Experiments are conducted on the public dataset LS-SSDD-v1.0. Experimental results prove our STDet has good performance for ship detection in large-scale SAR images.