PSPNet: Pretraining and Self-Supervised Fine-Tuning-Based Prototypical Network for Radar Active Deception Jamming Recognition With Few Shots

Abstract

To address the problems of requiring a large number of labeled training jamming samples in practical application, we propose the PSPNet, an improved prototypical network with pretraining and self-supervised fine-tuning, which can achieve high-precision radar active deception jamming recognition with few shots. For the time-frequency images of radar returns including deception jamming, we construct a deep network with multiple convolutional layers as the encoder. The encoder firstly preformes supervised pretraining on the Omniglot dataset with multiple categories, to extract the geometrical characteristics through learning. After that, for domain adaptation, the encoder is fine tuned in a self-supervised paradigm on the simulated unlabeled active deception jamming data. Finally, the encoder using prototypical network is applied to active deception jamming recognition under very few samples. Comparative experiments show that, the proposed PSPNet achieves high accuracy with 97.02% and 98.49% respectively in 5-shot case on both simulation and real data, which outperforms existing methods. The ablation experiments also demonstrate that both pretraining and self-supervised fine-tuning can improve the proposed network’s performance.