An Improved Oriented Ship Detection Method in High-resolution SAR Image Based on YOLOv5

Abstract

With the continuous development of space borne synthetic aperture radar imaging technology, massive volumes of high-resolution earth observation images have been used in marine ship monitoring. Aiming at the ship target with large aspect ratio, arbitrary direction and dense arrangement in high-resolution image, an improved oriented ship detection method based onYOLOv5 is proposed in this paper. First, considering the accuracy and effectiveness of model training, we recalibrated the SAR Ship Detection Dataset (SSDD) based on the oriented bounding box, which effectively improved the detection performance of deep learning-based detectors. Secondly, the mosaic method is used to enhance the data set and the loss function is modified to improve the model performance. In addition, the rotation detection algorithm with Circular Smooth Label (CSL) is applied to the YOLOv5 detection network, achieving precise positioning of ship target. Experiments on the recalibrated SSDD show that, compared with the classic deep learning-based detection method, the method proposed in this paper has obvious advantages in detection accuracy and detection efficiency.