MDGAN-TPAMDPN-Net: A Self-Sufficient and Calibrated Network for Radar Emitter Identification

This article develops the maximum diversity generative adversarial network (MDGAN)-TPAMDPN-Net for radar emitter identification (REI). 1) The number of training samples in the database is limited in real world, which leads to the inability to effectively support the training of the REI model. Thus, generative adversarial network (GAN) is designed to generate new samples to expand the training data set. 2) Since GAN is prone to generate samples with a single pattern, resulting in the poor generalization of the REI model. Herein MDGAN is proposed to increase the sample diversity. 3) There are some defects in traditional manual feature selection, such as low timeliness and low discrimination, so dual path network (DPN) is utilized to automatically extract and identify features. 4) In the process of feature extraction, the features may loss due to the deep layer of DPN, so we propose to add an attention module, triple path attention module (TPAM), to calibrate and resample the radar emitter feature maps, and embed it into the DPN to form TPAMDPN. 5) Finally, MDGAN and TPAMDPN are defined as the generative model and discriminative model, respectively, to form a complete REI network: MDGAN-TPAMDPN-Net. The results show that MDGAN-TPAMDPN-Net can self-generate samples to meet the training requirements under the condition of limited sample number, and improve the accuracy through feature calibration resampling.