Few-Shot Learning Method for Space Non-Cooperative Target Recognition

The recognition of space objects is crucial for tasks like non-cooperative rendezvous and docking, involving the identification of satellites, spacecraft, and space debris. This technology is essential for planning space activities, maintaining orbits, and avoiding collisions. While deep learning has advanced space object identification, these models struggle with rapid learning, especially when training samples are scarce due to high costs. This paper addresses these limitations by focusing on few-shot learning techniques for space object identification. We start with coarse-grained (category-level) identification, using knowledge extraction and fusion to enhance learning from minimal data. For fine-grained (model-level) classification, we develop multi-granularity techniques to accurately distinguish between different space object models. Additionally, we address knowledge degradation in continuous incremental learning by integrating it with few-shot learning, ensuring the maintenance and expansion of a persistent target library. Experiments on datasets like Mini-ImageNet and NWPU-RESISC45 validate our methods, demonstrating higher accuracy compared to traditional techniques. Our contributions improve rapid learning capabilities, accuracy, and continuous knowledge integration for space object identification, supporting safer and more efficient space missions.