View article

Synthetic aperture radar (SAR) has unique advantages in ocean monitoring. Ship object detection in multitemporal SAR images has great potentials in various applications. In this study, we aim to improve the accuracy of deep learning method for ship detection in SAR images. We have proposed a framework based on the Mask R-CNN. The training process utilizes SAR dataset SSDD, and the model incorporates ImageNet weights which are adequately trained for transfer learning. Experiments based on both Single-Mask and Multiple-Masks have been designed and conducted with data preparing methods such as denoising and data augmentation. The results showed that Mask R-CNN model outperformed other start-of-art models for SAR image-based ship detection. Specifically, the accuracy has been improved to 96.8% within 116ms.