Optical Remote Sensing Ship Object Detector Based on Improved GWD

In recent years, with the continuous development of remote sensing ship object detection technology, object detection technology based on the horizontal bounding box has made huge progress. However, in the optical remote sensing ship image dataset, the ship orientations are different and densely arranged, thus the horizontal bounding box is not suitable for this situation. Therefore, in order to solve the problem that the horizontal bounding boxes cannot accurately detect rotated ships, this paper proposes a remote sensing ship object detector based on IGWD (an improved Gaussian Wasserstain distance network), which adopts the data augmentation strategy of Cutout to simulate the situation where ships are occluded by clouds and fog, and improve the detection robustness of the overall network. To give a differentiable mathematical model to serve for BP algorithm for the positioning of rotated rectangle boxes, a loss model GWD (Gaussian Wasserstain distance) based on the Gaussian Wasserstain distance is used to simulate the deviation of the positions of two rotated rectangle boxes. In addition, this paper also introduces deformable convolutional networks in the backbone for arbitrarily oriented ships detection. Many ablation studies show that our method can achieve excellent performance on the famous ship dataset HRSC2016.