Few-shot Radar Jamming Recognition Network via Complete Information Mining

Effective recognition of radar jamming is of great importance in improving radar system's anti-jamming capability. Existing radar jamming recognition methods based on convolutional neural networks suffer from limitations such as low recognition accuracy and poor network stability in complex electromagnetic environments. This is due to incomplete extraction and utilization of useful information from the jamming signal. In this paper, a few-shot adaptive confidence aggregation and cross-modal refinement jamming recognition method (JR-ACAR) is designed to mine the complete information of the jamming signal for enhancing both accuracy and robustness of the method. Adaptive confidence aggregation (ACA) model is proposed to fully exploit the complementary information between the modulus, phase, real and imaginary parts of the input time-frequency spectrum of jamming signal, and adaptively aggregate these information by confidence vectors. Moreover, to further improve the recognition accuracy while enhancing the robustness of the network, a cross-modal refinement model (CMR) is proposed to mine the correlation information between the aggregated jamming recognition result and the original input time-frequency image. Experimental results on both simulated and measured mixed datasets validate that the proposed JR-ACAR is superior to existing methods for radar jamming recognition in terms of recognition accuracy and robustness.