ABSTRACT Low probability of intercept (LPI) radar signal recognition under low signal‐to‐noise ratio (SNR) remains a formidable challenge, as intercepted signals are often submerged in severe noise, leading to modulation feature ambiguity. Although generative models offer potential for signal restoration, applying them blindly can result in ‘hallucinations’ or incorrect structural guesses. To address these issues, this paper proposes a unified framework that integrates SNR‐guided generative restoration with an active recurrent glance‐and‐focus classification strategy. First, we introduce an RA‐UnetJiT restoration network, where the generative process is dynamically modulated by the input SNR via adaptive layer normalisation (AdaLN). This mechanism leverages U‐Net structural priors to strictly constrain the generation, ensuring the recovered spectrograms maintain semantic fidelity even at extremely low SNRs. Second, to address the computational redundancy of sparse spectrograms, we formulate feature extraction as a sequential decision process (SDP). A deep reinforcement learning (PPO) agent actively scans the restored image, selecting a sequence of informative high‐resolution glimpses. Finally, a transformer encoder models the spatiotemporal dependencies of these glimpses for robust classification. Experimental results on 16 types of radar emitters demonstrate that the proposed method effectively mitigates noise interference and achieves state‐of‐the‐art accuracy (e.g., at −10 dB) with significantly reduced computational cost compared to traditional sliding‐window baselines.
Cheng et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: