Learning ship width and direction by convolutional neural networks without manual labelling

摘要

Direction and width prediction have become hot topics in accurate ship location, while the resolution of remote sensing data has been increasing. In the current machine learning, demand for extensive manual labelling has become one of the biggest challenges to further improve tasks such as ship attribute prediction. To cope with this problem, in this paper, we propose a method to predict ship direction and width without manual labelling. To this end, pseudo-label generation approaches have been proposed for transfer learning with convolutional neural network (CNN). Experiments demonstrated that the pre-trained classification CNN features could preserve variational information for such attribute prediction. And the proposed pseudo-labels, which are even with limited qualities, could be efficient to train an accurate ship direction and width prediction model through fault-tolerant training by neural networks.