REMI: Few-Shot ISAR Target Classification Via Robust Embedding and Manifold Inference

Area of Science:

• Computer Science
• Artificial Intelligence
• Signal Processing

Background:

• Inverse Synthetic Aperture Radar (ISAR) target classification faces challenges from unknown image deformations and limited data (few-shot learning).
• Existing methods struggle with robust feature representation and accurate correlation modeling under these conditions.

Purpose of the Study:

• To propose a novel two-stage few-shot ISAR classification network, Robust Embedding and Manifold Inference (REMI), for enhanced target classification.
• To improve robustness against unknown image deformations and address few-shot learning limitations in ISAR target recognition.

Main Methods:

• A two-stage network: Robust Embedding Stage (Multihead Spatial Transformation Network - MH-STN, Grouped Embedding Network - GEN) and Manifold Inference Stage (Masked Gaussian Graph Attention Network - MG-GAT).
• MH-STN adjusts image deformations; GEN integrates and compresses features.
• MG-GAT captures sample manifolds using Gaussian-distributed node features and masked attention.

Main Results:

• REMI significantly enhances few-shot classification performance on ISAR datasets.
• The proposed network demonstrates robustness across various challenging scenarios.

Conclusions:

• The REMI network offers a robust solution for few-shot ISAR target classification, effectively handling image deformations.
• The combination of advanced embedding techniques and manifold inference provides superior performance and reliability.