Attention-based Semantic Segmentation of Satellites for In-orbit Space Situational Awareness

Edge-AI capabilities for on-board processing in space-based platforms has emerged as an important research field. Semantic segmentation, an important vision-based AI technique can assist with tasks like in-orbit servicing and space debris removal missions. Challenges for in-orbit semantic segmentation include structural variations in various spacecrafts, differences in patterns/textures for body and solar panels, low-lighting leading to significant SNR requirements, direct solar glare and complex background. To address these issues, we propose MANet for performing segmentation with MiT-based backbone. The network achieves state-of-the-art results for SPARK2024 dataset i.e. IoU scores of 0.9977 and 0.9974 for training and validation sets respectively.