Small ship detection based on YOLOX and modified Gaussian Wasserstein distance in SAR images

Due to the increase in data quantity, ship detection in Synthetic Aperture Radar (SAR) images has attracted numerous studies. As most ship targets are small and cover a few pixels in SAR images, the commonly used intersection-over-union (IoU) metric which is sensitive to the location deviation of the bounding box is not suitable to measure the distance between two small ship boxes. To solve this problem, this paper proposes a small ships-oriented detection method based on YOLOX. First, as an anchor-free one-stage detector, YOLOX can achieve state-of-the-art performance without extra anchor parameters. To make a balance between detection accuracy and speed, YOLOX-tiny is adopted as the baseline network. Then, a modified Gaussian Wasserstein distance is proposed. By modeling the bounding boxes as 2D Gaussian distributions, the Modified Wasserstein Distance (MWD) can be used to measure the similarity between the boxes in network training and post-processing. Finally, the proposed method is verified on Large-Scale SAR Ship Detection Dataset-v1.0 (LS-SSDD-v1.0), and the experimental results show that the proposed MWD can effectively improve the detection performance on small ships.