Few-Shot Semantic Segmentation Method for Crater Detection on Small Bodies

Craters are the primary landmarks used for visual navigation in missions exploring small celestial bodies. However, obtaining high-quality, annotated crater data is often challenging due to limited imaging conditions and strict mission constraints. Conventional semantic segmentation models struggle with limited data and are challenging to train effectively. To overcome this limitation, this study introduces a few-shot segmentation approach for crater detection on small celestial bodies. Our method includes a prototype representation module that constructs class-level prototypes to quickly associate crater regions with their semantic features. This paper also designs an iterative learning module that gradually improves the segmentation output, helping the model better capture detailed edges and structures. Tests on a simulated few-shot dataset demonstrate that our method provides reliable and accurate crater segmentation, achieving a mean intersection-over-union (mIoU) of 88.7, outperforming traditional fully supervised methods.