多域特征引导的无监督SAR图像舰船检测方法

如何在合成孔径雷达(SAR)图像标注样本有限的条件下，提升舰船检测性能一直是SAR图像处理中的热点问题。本文提出一种多域特征引导的无监督域适应方法，将知识从有标注的源域(光学图像)转移到未标注的目标域(SAR图像)，降低对标记SAR图像数据依赖。同时，设计了频域转换模块、注意力区域增强模块和自适应权重模块来缩小光学、SAR图像域之间的域差距，提高源域与目标域特征对齐效率，增强网络在挑战性样本下的特征迁移能力。在公开发布的数据集上进行了大量实验。结果表明:所提的模块较基础模型AP50提升10%，总体性能优于其他先进的方法。

A Multi-domain Feature-guided Method for Unsupervised Ship Detection in SAR Images

How to improve the ship detection performance with limited annotation samples in a synthetic aperture radar (SAR) image has always been a hot spot in SAR image processing.In this paper，a multi-domain feature-guided unsupervised domain adaptation method is proposed.The knowledge is transferred from the annotated source domain (optical images) to the unannotated target domain (SAR images)，and thus the dependency on the labeled SAR images is reduced.At the same time，the frequency domain transfer module，attention area-enhanced (AAE) module，and adaptive weighted module are designed to narrow the domain gap between the optical and SAR image domains，improve the efficiency of feature alignment between the source and target domains，and enhance the capability of feature transfer under challenging samples.Extensive experiments are carried out on public published datasets.The results show that the proposed modules are 10% better than the baseline，and the overall network outperforms other state-of-the-art (SOTA) methods.