Structure-Guided Domain-Adaptive Network for Few-Shot SAR Ship Detection.

Abstract

Due to the complexity of synthetic aperture radar (SAR) imaging mechanisms, SAR ship detection faces challenges such as difficulty in sample annotation and the influence of complex backgrounds, leading to poor target readability and difficulties in feature representation. Compared to SAR images, optical remote sensing images offer advantages such as high resolution and intuitive visualization, making them complementary to SAR images. Based on this, this article leverages the more intuitive structural features of optical remote sensing images to guide SAR ship target feature learning and proposes the structure-guided domain-adaptive network (SGDANet) for few-shot SAR ship detection. This proposed network follows a convolutional neural networks (CNNs)-transformer architecture, modeling the structural and edge token features of optical remote sensing images and embedding them into the network. Additionally, a feature fusion mechanism based on the split-fuse-merge strategy and attention mechanism is designed to achieve improved domain adaptation performance during the adversarial learning stage. Experiments on three self-built ship datasets illustrate that SGDANet outperforms other related network models in both three-shot and five-shot scenarios. Notably, SGDANet also exhibits good performance in zero-shot SAR target detection tasks, indicating its strong generalization ability. Code is available at https://github.com/RSIP-NJUPT/SGDANet.