RDYOLOv5m6-KF: A Rotation Detector for Ship Detection in Remote Sensing Images

Abstract

The use of remote sensing images for ship detection can accurately monitor ship targets and provide reliable reference for monitoring key sea areas. Since the horizontal detection model cannot precisely locate and represent the specific direction of the ship, we propose a rotation detector based on YOLOv5m6 and KFIoU, which can realize the detection of ships in arbitrary orientations. On the other hand, the punishment based on Gaussian Wasserstein distance is used in model to generate confidence loss, which improves the discrimination between foreground and background during ship detection. Finally, transformer pyramid attention is added to the backbone of network, which uses the fusion of information extracted in multi-scale space and the self-attention mechanism to improve the feature extraction effect and the accuracy of detection. On FGSD2021 dataset, our model finally achieves 88.24% of mAP after adding attention mechanism and improving the confidence loss.