Rotation Awareness Based Self-Supervised Learning for SAR Target Recognition With Limited Training Samples.

Area of Science:

• Computer Science
• Electrical Engineering
• Remote Sensing

Background:

• Synthetic Aperture Radar (SAR) automatic target recognition (ATR) is challenged by target pose variations, increasing the need for extensive training data.
• Existing SAR ATR algorithms struggle with limited sample scenarios due to sensitivity to azimuth angles.

Purpose of the Study:

• To develop a rotation-aware learning framework (RotANet) for SAR ATR.
• To address the challenge of limited training samples in SAR ATR by improving feature generalization.

Main Methods:

• Proposed an encoding scheme to characterize intra-class target rotational patterns.
• Developed a self-supervised task for RotANet to learn and generalize rotation awareness without external supervision.
• Integrated rotation awareness as a prior to regularize feature learning for the target recognition task.

Main Results:

• RotANet demonstrated significant improvements in recognition accuracy, particularly with very limited training samples.
• The framework achieved state-of-the-art performance compared to existing SAR ATR algorithms.
• Effectiveness validated through extensive experiments on a benchmark database for moving and stationary target recognition.

Conclusions:

• RotANet offers an effective solution for SAR ATR under limited training data conditions.
• The proposed rotation awareness mechanism enhances feature generalization and recognition performance.
• The self-supervised approach enables human-like rotation understanding, advancing SAR ATR capabilities.