MFE-Net: A Novel Multiscale Feature Enhancement Network for SAR Ship Instance Segmentation

Deep learning has achieved significant progress in ship instance segmentation for synthetic aperture radar (SAR) images. However, due to the challenges posed by inshore scenes, such as dense clustering, arbitrary arrangement orientation, and multiscale targets, the performance of existing methods still falls short of expectations. To overcome the dilemma, we propose MFE-Net, a one-stage instance segmentation network based on oriented bounding box (OBB). First, we design an Orientation-aware Segmentation Head (OSeg Head) to enhance the network’s perception of directional features while mitigating interference from complex background. Additionally, a Feature Aggregation Diffusion Pyramid (FADP) is proposed, consisting of the Multi level Feature Alignment Module (MAM) and the Multiscale Feature Enhancement Module (MFEM). The FADP aggregates global semantic and detailed information, and effectively enhances the feature representations of both deep and shallow layers through multiscale feature extraction. Experimental results on two public benchmark datasets (HRSID, SSDD) and a self-built dataset (HRSID_IS) demonstrate that MFE-Net achieves an ideal trade-off between accuracy and efficiency, respectively reaching 58.9%, 63.0%, and 46.9% AP with parameters of 12.36 M.