A Multiscale Hybrid Perception Network With Granularity Decoupling and Spectral Enhancement for HRRP Target Recognition

High-resolution range profile (HRRP)-based radar automatic target recognition (RATR) is crucial in capturing target structural characteristics across all-day and all-weather environments. However, existing HRRP-based RATR methods struggle to adapt to diverse target aspects and backscattering characteristics due to insufficient modeling of spatial structures and spectral periodicity. To address these limitations, we propose the MSHP-Net, a multiscale hybrid perception network with granularity decoupling and spectral enhancement for HRRP target recognition. The MSHP-Net employs a multiscale hybrid perception (HP) encoder to jointly capture spatial–spectral domain features by combining spatial feature decoupling and spectral feature recalibration. Specifically, we extract multigranularity spatial features and decouple them into granularity-invariant and granularity-variant components based on an adaptive singular value decomposition (SVD). This process explicitly enhances structural consistency and preserves fine-grained variations, effectively modeling varying target aspects and mitigating structural distortions inherent in HRRP data. To capture global spectral correlations and periodic scattering characteristics, we recalibrate the spectral distribution and apply spectrally enhanced attention, emphasizing the critical spectral bands and suppressing background noise. To promote multiscale hybrid features interaction, we introduce a hierarchical affinity-guided gating to propagate cross-scale relevant information flow, balancing low-level details with high-level semantics for more comprehensive feature representations. Finally, we aggregate scale-wise features and predict the final classification. Comparative experiments on both simulated and measured datasets validate the effectiveness of the proposed network.