Compact Polarimetric ISAR Space Target Components Recognition With Dual-Branch Correlation Aggregation Graph Attention Network

Inverse synthetic aperture radar (ISAR) enables continuous space surveillance irrespective of weather conditions, while the compact polarization (CP) mode balances hardware costs with the provision of polarization information. Recognizing components of space targets provides valuable insights into attitude inversion and the detection of abnormal motion. However, existing ISAR space target recognition methods lack the capability to transition from target-level classification to component-level recognition. Additionally, components with weak and nonuniform scattering intensity distributions pose challenges to their precise positioning and classification in ISAR images. To address these limitations, a coarse-to-fine dual-branch correlation aggregation graph attention network (DCA-Net) is proposed, featuring a novel graph neural network (GNN) construction strategy. Considering target sparsity, a multichannel graph node filter (MGNF) module combining compact polarimetric features is devised to enhance computation efficiency. Subsequently, a dual-branch correlation aggregation graph (DCAG) module is constructed concerning the local correlation and global topology. Information suppressed by the nonmaximum suppression (NMS) algorithm is reutilized to construct a local graph, which is then aggregated to improve the recognition probability of weak scattering intensity components. Meanwhile, a global graph is constructed, allowing the utilization of structural topology relationships for robust inference in heterogeneous scattering scenarios. The experimental results on the ISAR dataset demonstrate that the proposed method achieves superior performance, with at least a 5.38% improvement in the $F1$ score index and 5.21% improvement in the mean average precision (mAP) index.