SAR ship detection method in complex scenes based on enhanced VarifocalNet

Aiming at the problem that ship targets in synthetic aperture radar (SAR) images in challenging environments are easily affected by background clutter, resulting in false positives and missed detections, in SAR image ship detection.a SAR ship detection method in challenging environments based on improved VarifocalNet is proposed. Firstly, ConvNeXt is incorporated as the core network to enrich the receptive field of the model to improve the feature extraction ability of the network of multiple sized ships in complex backgrounds. Secondly, replacing the nearest neighbor interpolation upsampling module with the flow alignment module can compensate for the challenge of feature alignment within the feature pyramid and bolstering multi-scale feature fusion proficiency. Finally, introducing MPDIoU loss function is designed to expedite model convergence and refine target localization precision. The empirical outcomes demonstrate that the detection accuracy of the refined model on the SSDD and SAR-Ship-Dataset datasets has been enhanced by 4.3% and 5.9%, respectively, in comparison to the original VarifocalNet model. This substantiates that the algorithm can perform high-quality ship detection on SAR complex background images.