META LEARNING-BASED APPROACH FOR FEW-SHOT TARGET RECOGNITION IN ISAR IMAGES

Rapidly evolving deep learning methods have yielded remarkable performance in Inverse Synthetic Aperture Radar (ISAR) target recognition. However, training deep neural networks often requires large-scale annotated datasets. Due to the scarcity of ISAR images, it is challenging to obtain sufficient well-labeled ISAR datasets. Therefore, this paper considers Few-Shot scenarios and investigates the fast learning and generalization of the model via a Meta-Learning framework. The simulated experimental results illustrate that the Meta-Learning model presented in this paper outperforms traditional Machine Learning method K-Nearest Neighbor (KNN) in terms of testing accuracy, achieving a 72.79% improvement in 5-way 6-shot tasks. In addition, we propose Learning Gain as a criterion to measure the learning ability of the model.