SpaceSeg: A High-Precision Intelligent Perception Segmentation Method for Multi-Spacecraft On-Orbit Targets

Area of Science:

• Computer Vision
• Aerospace Engineering
• Machine Learning

Background:

• Accurate segmentation of multiple on-orbit spacecraft in deep-space imagery is challenging due to large uniform backgrounds, fine details, and limited labeled data.
• Existing methods struggle with the complexities of space imaging, hindering effective spacecraft monitoring and analysis.

Purpose of the Study:

• To develop an advanced segmentation framework, SpaceSeg, for accurate multi-spacecraft identification in deep-space imagery.
• To improve the robustness and efficiency of spacecraft segmentation models in challenging space environments.

Main Methods:

• SpaceSeg adapts a vision foundation model using a Multi-Scale Hierarchical Attention Refinement Decoder (MSHARD) and Spatial Domain Adaptation Transform (SDAT) training.
• A connected-component-analysis module is integrated for instance-aware organization in multi-spacecraft scenes.
• A new dataset, SpaceES, was created for multi-scale, on-orbit, multi-spacecraft semantic segmentation.

Main Results:

• SpaceSeg achieved state-of-the-art performance on the SpaceES dataset with 89.87% mIoU and 99.98% mAcc.
• The framework demonstrated superior accuracy compared to baselines, with 1.38 percentage points higher mIoU than the strongest competitor.
• SpaceSeg significantly reduced parameter count by 59.6% compared to a competing method and outperformed the vanilla SAM2 baseline by 5.71 percentage points.

Conclusions:

• SpaceSeg offers a highly accurate and efficient solution for multi-spacecraft segmentation in challenging deep-space imagery.
• The framework's robustness and performance were validated through hardware-in-the-loop simulations and real satellite imagery.
• The publicly available dataset and code will facilitate further research and development in spaceborne computer vision.