Synthetic pseudo-text aided radar interference recognition: a few-shot learning approach

In practical scenarios of radar interference recognition, it is challenging to acquire sufficient instances of certain interference patterns, yielding a limited dataset for training recognition models. This scarcity of training data severely degrades the recognition accuracy. To address this few-shot recognition dilemma, this paper proposes a novel few-shot interference recognition methods based on synthetic pseudo text dataset. Large Language Models (LLM) are adopted to generate varied text descriptions as supporting data. After splitting and filtering, the text data is paired with the time frequency (TF) graph of interference, forming an image-to-text dataset. We select the Contrastive Language-Image Pretraining (CLIP) model as the base model and further refine it using the synthesized text-image pairs. This fine-tuning process enables the model to recognize radar interference types by calculating the similarity between the signal’s TF graph and the descriptive text prompts for each potential class. Simulation experiments demonstrate that our proposed method achieves a recognition accuracy of 93.32% with a only 5 shots supporting set, outperforming ProtoNet and MatchingNet by 4.15% and 5.94%, respectively.