Learning a robust CNN-based rotation insensitive model for ship detection in VHR remote sensing images

Abstract

ABSTRACT Deep convolutional neural networks (CNN) have been widely applied in various fields, especially in the field of object detection. Deep CNN-based models showed great advantages over many traditional methods, even so, there are still many specific problems in the application of certain scenarios. In very high resolution (VHR) remote-sensing image datasets, the uncertainty of the object direction angle causes big trouble to the learning of the detector. Although the pooling operation can slightly alleviate the deviation caused by small angle, the feature learning of the objects with larger angle rotation still relies mainly on the sufficiency of sample data or effective data augmentation, which means the insufficiency of the training instances may cause serious performance degradation of the detector. In this paper, we propose a multi-angle box-based rotation insensitive object detection structure (MRI-CNN), which is an extended exploration for typical region-based CNN methods. On the one hand, we defined a set of directionally rotated bounding boxes before learning, and restricted the classification scene in a small angular range by rotated RoI (Region of Interest) pooling. On the other hand, we proposed a more effective screening method of bounding boxes, enabling the detector to adapt to diverse ground truth annotation methods and learn more accurate object localization. We trained our detector with different datasets containing different amount of training data, and the test results showed that the method proposed in this paper performs better than some mainstream detection methods when limited training data are provided in VHR remote-sensing datasets.