Spacecraft Detection Based on Deep Convolutional Neural Network

Spacecraft detection is one of essential issues on aerospace information processing and control, and can provide reliable dynamic state of target, so as to support decisions made on target recognition, classification, catalogue, et al. Although numerous spacecraft detection methods exist, most of them cannot achieve real-time detection, and are still lack of better accuracy and fault-tolerance for different scenes. Recently, deep learning algorithms have achieved fantastic detection performance in computer vision community, especially the regression-based convolutional neural network YOLOv2, which has good accuracy and speed, and outperforming other state-of-the-art detection methods. This paper for the first time applies CNN to the detection of spacecraft and sets up a dataset for target detection in space. Our method starts with image annotation and data augmentation, and then uses our improved regression-based convolutional neural network YOLOv2 to detect spacecraft in an image. The experimental results have shown that our algorithm achieves 97.8% detection rate in the test set, and the average detection time of each image is about 0.018s, which has lower time overhead and better robustness to rotation and illumination changes of spacecraft.