SSJR-Net: Semi-Supervised Few-Shot Radar Jamming Recognition Network

Deep learning-based radar jamming recognition focuses on identifying the jamming types by using the convolutional neural network (CNN). Though achieving superior performance in recent years, existing deep learning-based few-shot methods are still unable to break through the barrier of needing at least 5% labeled data for supervised learning, and fail to utilize the large amounts of unlabeled data. Considering the difficulty and formidable cost of acquiring labeled data and the ease of acquiring large amounts of unlabeled data in real scenarios, it is meaningful to explore a semi-supervised jamming recognition method by utilizing fewer labeled data and extensive unlabeled data. To this end, a few-shot semi-supervised radar jamming recognition network via a self-training framework is proposed in a challenging setting (1% labeled data, 5 labeled data for each jamming class). To effectively mine the recognition-related knowledge from the labeled data, a mutual learning strategy is first proposed by constraining the consistency between the predictions on original jamming data and their augmented data. With the trained mutual learning strategy, the pseudo-labels of unlabeled data can be obtained by taking the unlabeled jamming data as input. Then, to select more reliable pseudo-labeled data, a pseudo-labeled samples selection mechanism is proposed by introducing confidence scores to filter the high-quality pseudo labels. With the two above mentioned components, our framework is able to effectively exploit information contained in both labeled and unlabeled data through self-training in a semi-supervised manner. Extensive experiments on dataset with a mixture of simulated and measured data demonstrate that the proposed semi-supervised jamming recognition method (SSJR-Net) outperforms the state-of-the-art techniques in few-shot jamming recognition and semi-supervised recognition.