Arbitrary-oriented Ship Detction Based on R2CNN
Mingming Zhu, Hao Zhou, Guoping Hu, Zongxin Liu · 2023
To solve the problems of arbitrary orientation, large aspect ratio and dense distribution for ship detection in SAR images, an arbitrary-oriented ship detection method based on rotational region convolutional neural network (R2CNN) is proposed. Firstly, the region proposal network (RPN) is used to generate arbitrary-oriented bounding boxes. Secondly, the convolutional feature maps are input into several region of interest pooling layers (RoIPoolings) with different pool sizes and the pooled features are linked for predicting the ship/non-ship scores and bounding boxes. Finally, the detection results are got by inclined non-maximum suppression (NMS). The experimental results show that the proposed method can effectively detect arbitrary-oriented ships and is superior to other arbitrary-oriented object detection methods.