Arbitrary-Oriented Ship Detection Based on Rotation Region Locating Networks in Large Scale Remote Sensing Images

Although deep learning has dominated the ship detection domain, it still has two challenges: arbitrary-oriented densely arranged ships cause detection omissions and large scale image contains redundant areas. This paper proposes an effective convolutional neural network framework for arbitrary-oriented ship detection in large scale and complex scenes. In this framework, we propose Cumulative Feature Pyramid Networks for multi-receptive-field feature fusion, which enhances high-level semantic information at all scales. Based on the outputs of Region Proposal Networks, Rotation Region Locating Network predicts rotation bounding box of arbitrary-oriented ships and adopts rotation intersection over union to avoid the effect of ship dense arrangement. For large scale scenes, No-Ship Area Suppression uses OTSU algorithm to generate binary mask and then filter out non-ship regions to reduce redundant computations. Additionally, we firstly build a remote sensing image dataset for ship detection, which contains 4237 images and over 21200 ships. Experimental comparisons with the state-of-the-art approaches validate the effectiveness, accuracy and robustness of the proposed method.