Radar intra-pulse signal modulation classification is a key problem in the radar countermeasure field, and the deep learning networks are more and more used to improve the classification performance. However, most of them rely on a large amount of labeled data. To solve this problem, a semi-supervised learning (SSL) network with the pseudo-labeling and consistency regularization methods (CRPL-SSL) is proposed, which can classify the complex modulated signals with limited labeled and unlabeled samples. The CRPL-SSL is constructed by a two-channel framework to augment the unlabeled data with weak and strong augmentation strategies, separately. The weak augmented samples are labeled with pseudo-labels, while the prediction probabilities are generated with the strong augmented samples. Then the pseudo-labels and prediction probabilities are used to refine the final labels by the consistency regularization method, in which the adaptive threshold is used to increase the reliability of the label judgement. The performance of the proposed CRPL-SSL model is compared with that of the other supervised learning (SL) models and SSL ones, which indicates the CRPL-SSL achieving a better classification accuracy, especially in the situations of limited number of samples and lower signal-to-noise ratios (SNRs).
Cai et al. (Sun,) studied this question.