Аңдатпа
Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR) remains challenging due to speckle noise, aspect-angle variation, and the loss of fine scattering cues in conven-tional deep-learning pipelines. Spatial-domain CNNs primarily extract geometric struc-ture but overlook high-frequency information critical for distinguishing small or spectrally similar targets, while frequency-only methods such as FFTNet fail to leverage spatial con-text and multi-scale spectral variation. To address these limitations, this study proposes the Multi-Scale Spectrum Pyramid Network (MSP-Net), which decomposes SAR images into low-, mid-, and high-frequency components via two-dimensional Fourier transforms with band-pass filtering and processes each band through dual convolutional branches equipped with predefined and learnable spectral filters. The resulting features are fused using attention-based, MLP-based, or transformer-based integration mechanisms. Expe-riments on two MSTAR-based benchmark datasets (11-class and 8-class) demonstrate that MSP-Net substantially outperforms spatial-only CNNs and single-scale frequen-cy-domain models. In the 11-class setting, MSP-Net improves accuracy by 13–14% (up to 95%) and achieves near-perfect ROC separability (AUC ≈ 1.0) with reliable calibration (ECE < 0.02). On the reduced 8-class dataset, the best MSP-Net variant achieves 99.9% ac-curacy and consistent per-class F1-scores. Ablation studies confirm the critical role of multi-scale spectral decomposition and adaptive fusion in improving recognition of small and spectrally similar targets such as BMP2, BTR60, and BTR70. These results highlight the effectiveness of frequency-aware, multi-scale learning for robust and interpretable SAR ATR.