An Improved FCOS Method for Ship Detection in SAR Images

More convolutional neural network (CNN) methods are widely utilized for ship detection in synthetic aperture radar (SAR) images. Nevertheless, there are still some problems that need to be addressed. First, detection methods may misclassify islands and onshore objects as ships. Second, due to the large scale and low resolution, a majority of ships are regarded as small targets and difficult to detect. Third, due to the different sizes of ships, the hyperparameters of anchor-based methods are difficult to set. Thus, we propose an anchor-free method called improved fully convolutional one-stage object detector (Improved-FCOS) to solve these problems. First, a multilevel feature attention mechanism is proposed to extract effective features and collect global context information. Second, we propose a feature refinement and reuse module with two stages to refine small ship features. Third, a head improvement module is designed to optimize the methods for classification and localization of ship targets. Finally, a modified varifocal loss is adopted to better train the classification branch. We conduct ablation experiments on LS-SSDD-v1.0 and verify the detection performance of the proposed modules. Additionally, we compare the Improved-FCOS with other state-of-the-art methods by experiments on three datasets, including the SAR-Ship-Dataset, HRSID, and LS-SSDD-v1.0. Detection results show that Improved-FCOS obtains the best detection performance on all datasets and indicate that the Improved-FCOS is more accurate and robust.