Radar Jamming Recognition Method Based on Cross-Modal Multilevel Feature Fusion

Abstract

Effective radar jamming recognition is a critical precondition for enhancing radar antijamming capabilities. Although deep neural networks have been widely adopted for this task, existing methods mainly rely on time-frequency (TF) maps, overlooking inherent signal features such as amplitude and phase. This incomplete representation leads to a significant decline in recognition accuracy under low jamming-to-noise ratio (JNR) and complex interference conditions. To address these challenges, we propose a cross-modal multilevel feature fusion network (CM-FF), which innovatively integrates one-dimensional signal tensors, spectrum and two-dimensional TF images to compensate for information loss in single-modal approaches, significantly enhancing feature separability and identification accuracy. A multilevel feature extraction module is proposed to extract multiscale features from both one-dimensional (1D) tensors and two-dimensional (2D) images. Besides, a multimodal feature fusion module is proposed to assign weights to different features adaptively. Experimental results show that our proposed method achieves a recognition accuracy of 98.4%, representing a maximum improvement of 14.6% over existing methods. Even under extremely low JNR conditions of −10 dB, our network maintains an accuracy rate of 80.75%. Furthermore, the network has fewer than 1 million parameters, demonstrating its lightweight design and low resource requirements.