Cross-Domain Multi-Task Representation Learning for Target Recognition with Dynamic Attitudes

In the realm of radar automatic target recognition, the adoption of high-resolution range profiles is steadily gaining popularity. However, this research grapples with a pivotal challenge: the alteration of attitude angles significantly influences the accuracy of airborne target recognition. To address this issue, we introduce a novel self-supervised training method called Cross-Domain Multi-Task Representation Learning. This method enhances feature robustness by integrating frequency-domain auxiliary tasks with temporal representation learning tasks, effectively adapting to scenarios where there is a noteworthy shift in data distribution. The approach exhibits a substantial enhancement in generalization capabilities, particularly in the context of zero-shot and few-shot classification tasks, amidst significant variations in data distribution.