Semi-Supervised Structural Prior-Guided Network for Space Target Component Segmentation in ISAR Images

Area of Science:

• Space Situational Awareness
• Artificial Intelligence
• Computer Vision

Background:

• Component segmentation in Inverse Synthetic Aperture Radar (ISAR) images is crucial for space situational awareness.
• Existing deep learning models struggle due to limited annotated data, inter-class confusion, and lack of structural priors.

Purpose of the Study:

• To propose a Semi-Supervised Structural Prior-Guided Network (SSPNet) for improved ISAR image segmentation.
• To address data scarcity and enhance model generalization for space target components.

Main Methods:

• Introduced a Gated Manifold-Constrained Hyper-Connections Vision Transformer (GMHC-ViT) encoder for broader feature representation.
• Developed a Prior-Guided Module (PGM) to extract and inject structural shape and edge priors.
• Implemented a tailored perturbation strategy for consistency regularization using unlabeled ISAR data.

Main Results:

• SSPNet demonstrated superior performance compared to existing methods on a simulated ISAR dataset with 38 target classes.
• The network achieved strong segmentation capabilities even under low signal-to-noise ratio (SNR) conditions.
• Alleviated inter-class confusion and enhanced cross-category generalization through adaptive gating and cross-attention mechanisms.

Conclusions:

• SSPNet effectively leverages structural priors and semi-supervised learning for robust ISAR image segmentation.
• The proposed architecture shows significant potential for enhancing space situational awareness tasks.
• The method provides a viable solution for component segmentation challenges in ISAR imagery.