A Multi-Branch Network for SAR Ship Detection

Synthetic Aperture Radar (SAR) can obtain reliable radar images without considering weather conditions, and has important military value in the field of ocean detection and reconnaissance. However, the imaging principles and complex backgrounds results in significant noise in SAR images, posing substantial obstacles to accurate ship target detection. Additionally, the variability in the scale and shape of ship targets increases the risk of missed detections. Aims at the above problems, this paper proposes a Multi-Branch Network (MBNet) based on YOLOv5 to improve the performance of SAR ship detection. Firstly, to capture ship targets of varying sizes, we introduce a novel Multi-Branch Dilated Convolution Block (MBDCB) that integrates cascaded dilated convolutions with convolution layers of different kernel sizes. Secondly, we expand the receptive field by adding a new pooling branch and incorporate spatial and channel attention mechanisms to mitigate the interference from complex backgrounds. Finally, we design a Two-Branch Deformable Convolution Block (TBDCB) that combines deformable convolutions with channel attention mechanisms to enhance the feature representation of ship targets. We conducted simulations on the SSDD and HRSID datasets, with results demonstrating that the proposed method significantly outperforms advanced techniques in terms of AP50, accurately detecting multi -scale targets and distinguishing ships from the background, thereby validating its effectiveness and robustness.