FuseCLIP: Semantic-Guided Multidomain Fusion for Few-Shot Radar Active Jamming Recognition

Abstract

Radar active jamming recognition is an essential component of radar anti-jamming processing and cognitive radar decision making. As jamming categories proliferate and electromagnetic environments become more complex, collecting sufficient labeled samples for every possible jamming condition can be challenging in practical scenarios. Thus, it becomes necessary to exploit the recognition capability under few-shot conditions, especially in the face of varying jamming-to-noise ratio (JNR) and compound interference. Existing methods commonly rely on a single signal representation or treat jamming categories as discrete labels, which may underuse the complementary evidence available across signal domains and the semantic knowledge associated with jamming categories. This paper proposes FuseCLIP, a semantic-guided multidomain fusion framework for few-shot radar active jamming recognition. FuseCLIP jointly encodes time-domain sequences, frequency-domain sequences, and short-time Fourier transform (STFT) spectrograms to capture waveform, spectral, and time-varying spectral characteristics. The resulting features are adaptively fused and mapped into a unified representation space, where they are matched with fixed Contrastive Language–Image Pre-training (CLIP) text prototypes generated from task-related jamming descriptions. By combining multidomain physical information with category-level semantic priors, the proposed framework is intended to make more effective use of limited labeled data. Experiments on a simulated radar active jamming dataset under different shot numbers and JNR levels demonstrate the effectiveness of FuseCLIP relative to representative baselines. Ablation experiments further indicate the complementary contributions of the time-domain, frequency-domain, and time-frequency inputs.