Efficient Instance Segmentation Method For High-Resolution SAR Imagery

Abstract

Benefiting from the characteristic of synthetic aperture radar (SAR), ship detection plays a crucial role in military purposes recently. Meanwhile, the methods developed by convolutional neural networks are superior in precision, detection speed, and migration ability to traditional methods. However, bounding boxes used in ship detection merely perform instance-level detection which comprises ship and background in the detected result. In terms of this defect, we introduce instance segmentation in SAR imagery for extracting the contour of the ships. Compared to ship detection with bounding boxes, instance segmentation implements pixel-level prediction for the ships, and the predicted mask eliminates background information, which provides the contour and azimuth of the ships. Furthermore, we proposed DB loss and high-resolution representations for instance segmentation. Quantitative and visualized results demonstrate proposed method improves average precision (AP) and reduces the model size compared to Mask R-CNN. Besides, it can accurately segment the hard samples, e.g., offshore ships, small ships, and ships in the port, which has a broad application vision in interpreting SAR images.