A Factor-Ignored Module for Easy Detection of Uncooperative Spacecraft Using Small Training Samples

The domain gap in neural networks has been a major problem limiting the efficacy of real-world applications. Attempts have been made in much research to close the domain gap using a few target samples. However, for uncooperative spacecrafts, both source and target sets for practical scenarios are difficult to obtain. In this article, the domain gap is approximately decomposed into different influencing factors and they are individually investigated. Therefore, a simple module called the factor-ignored module is proposed, which aids neural network training to allow spacecraft information to be learned with less interference from other image information. Meanwhile, a comprehensive training method is formed based on this module by using YOLOv5 as an example framework with color space and Earth background factors. Only 600 synthetic images are utilized for training in order to avoid using numerous datasets. All networks are tested on the subset of SPEED and HOANG datasets in order to simulate uncooperative spacecraft conditions and get the test set closer to the real scenario. The experimental results prove the effectiveness of the module. Especially ignoring the influence of the Earth background can greatly improve the network detection precision and the color space mainly affects the recall.