Accurate ship segmentation via ship contour prediction

摘要

Accurate ship segmentation in optical remote sensing images is challenging. In this paper, ship contour prediction network is introduced to improve the performance of ship segmentation network. Ship contour prediction network predicts the ship contour, which can promote the learning of ship segmentation network, producing more accurate segmentation results. The ship contour prediction network can be naturally embedded into any ship segmentation network with encoder-decoder architecture. Moreover, ship contour prediction network is only used in the training process, and thus the network not only can improve segmentation accuracy, but also does not increase computational cost in the testing process.