Fusing Feature Pyramid and Self-Attention Mechanism for SAR 3D Point Cloud Target Recognition

‍ ‍Synthetic aperture radar （SAR） is increasingly recognized as a critical method for target recognition that can obtain the scattering feature image of the target. However， the two-dimensional （2D） image of traditional SAR has a problem of height direction occlusion effect， which seriously affects the accuracy of target recognition. Through multiple observations， SAR 3D imaging can form a synthetic aperture in the height dimension， improve the resolution， and distinguish overlapping targets， which is the future direction of SAR research. However， compared with SAR 2D images， SAR 3D images have critical features such as sparse point clouds and uneven distribution. The existing SAR 2D recognition methods find it difficult to effectively handle these features and have poor recognition accuracy. Existing SAR 3D recognition methods directly transfer the optical recognition network. However， owing to the characteristics of SAR 3D images， such as sparse point clouds and uneven scattering intensity distribution， the performance of SAR 3D recognition applications is degraded because the optical recognition network is mostly oriented to dense point clouds， and it is usually difficult to make full use of the scattering information. To address these problems， this paper introduces a SAR 3D point cloud target recognition neural network fusing feature pyramid and self-attention mechanism， which significantly enhances the accuracy of SAR 3D target recognition. Specifically， the proposed method uses the feature pyramid to construct and fuse multi-level features， which improves the utilization ability of shallow features and deep features of the point cloud， and solves the problem of information loss caused by the sparsity of the SAR 3D point cloud. The self-attention mechanism is employed to adaptively adjust the local semantic relationship of the target point cloud， enhance the network ability to extract the features of strong scattering areas， reduce the influence of the features of weak scattering areas， and solve the problem of target recognition accuracy degradation caused by the uneven distribution of SAR 3D point clouds. A microwave anechoic imaging system was developed， and eight types of ground vehicle target 3D data were collected to create data sets. Target recognition performance comparison experiments， ablation experiments， and feature dimension reduction visualization experiments were carried out. The experimental results demonstrate the advantages of the proposed method in terms of recognition accuracy.