Research On SAR Ship Detection Technology Based On Deep Learning

Synthetic Aperture Radar(SAR)is a high-resolution radar technology with all-day,all-weather,and long range.It can obtain high-precision target measurement near the ground and in the atmosphere.In the field of ocean observation,SAR has become an indispensable primary means.For the field of SAR ship object detection based on deep learning,the data set data quantity is small,low detection accuracy,the detection accuracy is not high and the horizontal detection box does not satisfy the demand of SAR ship detection.this paper proposes a SAR ship detection method based on self-supervised learning and a SAR ship detection method based on a rotated bounding box.The main research contents of this paper are as follows:(1)A self-supervised learning pre-trained SAR ship detection methodology using Swin Transformer is put forward for the current SAR ship detection dataset with low data amount and difficult manual labeling,which makes it difficult to extract features using large-scale backbone networks and leads to restricted detection accuracy.The method introduces the self-supervised learning method MoBY into the field of SAR ship detection,and can obtain higher detection accuracy than the traditional supervised method by using only the training set of SSDD dataset.Faster R-CNN using MoBY has an AP of 1.4%higher than supervised pretrain,compared to before the improvement,the average persion has increased by 1.9%.(2)Aiming at the current SAR ship detection method which is not optimized for high quality detection and the detection frame localization accuracy is not high,the positive and negative sample assignment and localization loss function are improved,and the DIoU R-CNN network which improves the Faster R-CNN is proposed in combination with the self-supervised learning method MoBY to realize the high accuracy SAR ship detection,achieved an average precision of 73.0% on the SSDD dataset.(3)Aiming at the inaccurate positioning of the SAR horizontal object detection method in complex scenes,and the large overlapping area of the horizontal detection frames in the case of dense coastal ships,a SAR ship rotation object detection method based on Rotated RepPoints is proposed.Different types of ships have different aspect ratios,and the proposed method improves the evaluation strategy of prediction points according to the object shape,incorporates the label assignment method KLDATSS,and proposes a new rotating object detection loss function KLDRepLoss for the limitations of existing rotating object detection representation methods.experiments show that the proposed method has higher.The proposed method is shown to have higher detection accuracy than other rotation detection methods,achieved an average precision of 79.9%on the rotation-labeled SSDD dataset.