Self-supervised SAR ship detection
Xiaokang Ren, Nannan Cai · Third International Conference on Computer Vision and Data Mining (ICCVDM 2022) · 2023
Synthetic Aperture Radar (SAR) has become one of the primary means of current earth observation due to its unique technical advantages, such as all-weather, all-time, and extended operating distance. However, most SAR detectors based on deep learning methods use outdated ResNet backbone networks, and the detection model detection accuracy is low. This paper proposes a new network called Dynamic IoU R-CNN (DIoU R-CNN) to transfer the self-supervised learning method moby based on Swin Transformer to the complex downstream task of SAR ship detection. DIoU R-CNN adds the dynamic IoU module to Faster R-CNN and the advanced BalancedL1 loss function, achieves relatively high accuracy SAR ship detection in the SSDD dataset without much increase in the number of parameters and training time. And the Swin Transformer with self-supervised learning performs even better than the supervised learning method in the comparison experiments.