Rotated SAR Ship Detection based on Gaussian Wasserstein Distance Loss

Deep learning-based rotated ship detection algorithms in Synthetic Aperture Radar images suffer from low detection accuracy and converge speed due to the boundary discontinuity and angle sensitivity problems. At the same time, in complex scenarios such as inshore, the detection accuracy is limited due to more interference. To address these problems, this paper proposes a rotated SAR ship detection algorithm based on the Gaussian Wasserstein Distance (GWD) loss function and salient feature extraction network. Based on the anchor-free detection framework, the rotated bounding boxes are converted to two-dimensional Gaussian encodings, and the Wasserstein distance between the distributions is used as the loss function of the rotated bounding boxes, and the model is guided to focus on the key features by the salient feature extraction network. Experimental results on the publicly available rotated SAR ship dataset SSDD+ demonstrate that the proposed method obtains remarkable performance compared to one-stage and anchor-free methods, especially that the average precision (AP) of the proposed method is 79.30% in the nearshore scenario, which is 4.90% higher than the suboptimal method.