SAR-ShipSwin: enhancing SAR ship detection with robustness in complex environment.

Contemporary synthetic aperture radar (SAR) image processing techniques face various challenges, particularly in ship detection, background noise reduction, and information preservation. To address these issues, this paper introduces a novel model we called SAR-ShipSwin, which combines the swin transformer and feature pyramid network as the backbone network structure, specifically designed for ship detection in SAR images. The backbone network optimizes computational efficiency and handles occlusion and overlap issues in SAR images successfully by introducing the improved window multi-head self-attention module. To further enhance recognition accuracy, we design the background modeling network, which efficiently identifies and eliminates complex background features. Additionally, we introduce the spatial intensity geometric pooling technique, a novel pooling strategy that preserves geometric and structural information of the original region of interest, significantly reducing information loss and distortion. Considering the diverse ship shapes in SAR images, we specially design the dynamic ship shape adaptive convolution module, which dynamically adjusts the shape of convolution kernels to better match the targets. The proposed model is validated on the SSDD and HRSID datasets, achieving state-of-the-art performance. [ABSTRACT FROM AUTHOR]