FS-YOLO: a multi-scale SAR ship detection network in complex scenes

Ship image recognition by synthetic aperture radar (SAR) is a crucial technology for intelligent shipping and maritime safety monitoring. However, existing ship detection algorithms are ineffective in detecting multi-scale ships in complex scenes. Therefore, we propose an FS-YOLO for detecting ship targets in complex backgrounds. Firstly, we design a feature enhancement module (FEM) using dilated convolution to enhance the feature-proposing ability of the backbone network. In addition, we propose a spatial channel pooling module combined with enhanced spatial pyramid pooling (ESPPCSPC) using channel attention to improve the expressive capability of the network. In addition, we achieve weighted feature fusion of different feature layers through a heightened bidirectional feature pyramid network structure (HPAN). Finally, we design a new confidence loss function to improve the robustness of the model to multi-scale targets. The experimental results show that the mAP50 values of FS-YOLO reach 96.9% and 89.5% on SSDD and HRSID datasets, respectively. Meanwhile, by comparing the performance of FS-YOLO with some commonly used detection algorithms, we discover that FS-YOLO has higher detection accuracy and good generalization.