A Three-Branch SAR Target-Recognition Network via Multi-Physical-Dimensional Collaboration

Abstract

Conventional deep-learning methods for synthetic aperture radar (SAR) target recognition predominantly rely on image-domain visual features, with insufficient integration of the physical principles governing SAR imaging. This limitation leads to unsatisfactory recognition accuracy and poor generalization performance under limited training-sample settings and complex operating conditions. To address this gap, we propose a three-branch collaborative network across multiple physical dimensions. First, the target orientation perception branch employs learnable anisotropic elliptical Gaussian convolution kernels to capture complete structural details. Second, the scattering characteristic modeling branch constructs topological structures aligned with the spatial distribution properties of scattering centers in a macro-to-micro progressive manner; it simulates the energy coupling process among scattering centers and mines global scattering attributes. Third, the azimuth–spatial evolution capture branch focuses on the azimuthal evolution regularity. By mapping SAR images into ordered token sequences, it extracts local azimuthal features and adopts a global query attention mechanism to model long-range dependencies. Experiments on two distinct SAR target datasets verify that the proposed model achieves up to 0.4% accuracy improvement compared with baseline models based on single-physical-feature extraction. Specifically, the model yields a maximum accuracy increment of 4% under extremely limited training samples and outperforms the optimal baseline by 0.15% under intense noise interference.