Robust Radar HRRP Recognition Under Nonuniform Jamming via a Frequency Attention and Reconstruction Fusion Network

Abstract

Radar automatic target recognition (ATR) based on high resolution range profile (HRRP) has witnessed rapid development in recent years. However, the HRRP recognition under electronic countermeasures (ECMs) remains a significant challenge. In particular, the jamming power received by the wideband radar system is always unevenly distributed across different subbands during the adversarial process of radar electronic jamming and countermeasures. Classical jamming suppression methods based on whitening filtering theory suffer from severe distortion of target HRRPs, rendering conventional target recognition methods ineffective. To address this problem, this article proposes a dual-branch deep learning architecture that combines frequency-domain adaptive filtering with time-domain signal reconstruction. Specifically, we introduce a complex-valued frequency attention (CFA) module acting as a learn-able spectral filter that reweights the spectrum according to the balance of jamming suppression and target HRRP distortion, motivated by the frequency-selective weighting concept from adaptive filtering. A parallel U-shaped network (U-Net) branch performs nonlinear reconstruction and extracts hierarchical bottleneck features for classification. The two branches are integrated through feature-level fusion in latent space, and a hybrid loss function balances reconstruction fidelity and classification accuracy to enable joint optimization. With the complementary contributions of both branches, the model learns robust representations invariant to diverse jamming distributions. Both simulated and measured-data experiments demonstrate that the proposed method consistently outperforms the state-of-the-art baselines, while exhibiting strong generalization capability under mismatched training-testing conditions.