A Novel Data Augmentation Method for Detection of Specific Aircraft in Remote Sensing RGB Images

In this paper, we propose a novel data augmentation method for the detection of specific aircraft in remote sensing RGB images. For object detection, the number of training data has great influence on the training of deep learning network. Researchers suggest that the extensive training samples in deep learning are indispensable. The extensive training samples can guarantee the accuracy and robustness of object detection. We refer to military aircraft and helicopter as specific aircraft. Due to the number of remote sensing images of specific aircraft is far less than that of civil aircraft, it is difficult to train an ideal detection model only by using the available specific aircraft images. Deep learning networks have the excellent ability on fault tolerance and generalization and can extract features from simulated aircraft samples. This means that the simulated aircraft samples can partly replace real images to some extent and reduce the need for detection models. Inspired by this, true remote sensing images are combined with specific aircraft three-dimensional models to form simulated images. Compared with previous data augmentation method, such as flipping and rotating, our method brings in new sample information. The experiments based on remote sensing images show the feasibility and effectiveness of our method. Meanwhile, the proposed method has compatibility with other data augmentation methods.