Group-Wise Feature Fusion R-CNN for Dual-Polarization SAR Ship Detection

Ship detection in synthetic aperture radar (SAR) images is a hot pot in the remote sensing (RS) field. However, most existing deep learning (DL)-based methods only focus on the single-polarization SAR ship detection without leveraging the rich dual-polarization SAR features, which poses a huge obstacle to the further model performance improvement. One problem for solution is how to fully excavate polarization characteristics using a convolution neural network (CNN). To address the above problem, we propose a novel group-wise feature fusion R-CNN (GWFF R-CNN) for dual-polarization SAR ship detection. Different from raw Faster R-CNN, GWFF R-CNN embeds a group-wise feature fusion module (GWFF module) into the subnetwork of Faster R-CNN, which enables group-wise feature fusion between polarization features and multi-scale ship features. Finally, the experiments on the dual-polarization SAR ship detection dataset (DSSDD) demonstrate that GWFF R-CNN can yield a ~4.1 F1 improvement and a ~2.9 average precision (AP) improvement, compared with Faster R-CNN.

I. Introduction

Ships, as the main carriers of marine activities, are of great value. As a main part of ship intelligent surveillance, ship detection plays an essential role in the remote sensing (RS) field. Due to the all-day and all-weather ability of synthetic aperture radar (SAR), ship detection in SAR images has attracted increasingly attention. Specifically, ship detection in SAR images has a huge potential in following applications, i.e., marine traffic monitoring, port vessel dispatching, marine sovereignty maintenance, maritime emergency rescue, marine fishery management and so on [1]–[4].