Rotationally Unconstrained Region Proposals for Ship Target Segmentation in Optical Remote Sensing

The segmentation of ship targets in remote sensing images is of great importance in both military and civil fields. The existing methods often suffer from a low overlap between region proposals, which are constrained to be the horizontal rectangles and the ground truth. This enhances the problems of background noise and results in more false positives and a lower recall. Hence, we propose a method to generate the rotated bounding boxes that are more appropriate for ship segmentation. The proposed method is based on the Mask R-CNN framework, and the key contribution lies in the approach to predefine the rotated bounding boxes and generate the rotationally unconstrained region proposals. Specifically, for a double-stage proposal generation, a multiangle region proposal layer is designed; following this, an adapted alignment approach is proposed to extract the features for each proposal. In addition to the above-mentioned two proposed methods, we adopt a second stage to refine the proposals with rotation regression and mask prediction. We evaluated the proposed method using a remote sensing dataset with extensively labeled ship targets, and the experimental results show that the proposed method performs better than its competitors. The proposed method significantly increases the recall of densely arranged ships while substantially reducing the number of false positives.