Robust Spacecraft Image Interest Point Detection and Description with Transformer Under Variable Illumination Conditions

The spacecraft image matching serves as the foundation for various space vision applications, such as vision-based relative pose estimation of target spacecraft. Currently, feature-based methods employ the extraction of local image features to compute interest points and descriptors for image matching. However, significant variations in spacecraft image local features occur when illumination conditions change, posing challenges for image feature matching. This paper proposes a Transformer-based self-supervised network for interesting point detection and description, which is trained under varying lighting conditions. The global receptive field provided by the Transformer enables the establishment of more robust feature descriptors, thereby allowing robust feature matching under varying lighting conditions. Additionally, a multi-illumination-angle spacecraft image dataset is created for training and testing the proposed algorithm. Compared to other state-of-the-art interesting point detection and description algorithms, the proposed algorithm achieves the best results in terms of matching effectiveness.