基于酉 ESPRIT 的多频带融合ISAR成像

基于散射中心参数化模型和反向传播（back propagation, BP）神经网络，构建了一种针对目标全角度、宽频段下的远场电场预测网络，该网络将利用目标的位置、幅度、频率等数据信息实现远场电场实部与虚部的快速预测. 首先，将对目标强散射点的位置以及强度等参数进行提取；然后，对二维角域以及频域进行区域划分，构建并联式的智能网络架构，从而建立散射中心参数化模型与高精度远场电场间的关系. 该方法能够通过新型并联网络的训练，减小传统散射中心模型的频率、角度依赖性的影响，实现目标远场电场的快速获取. 由于在网络设计时，充分借鉴了现有的模型中各散射参数对目标电场的影响，因此该神经网络具有清晰的物理意义以及突出的泛化能力. 与传统的基于几何绕射理论（geometrical theory of diffraction, GTD）模型的电场重构方法相比，本文方法具有更高的准确性，实验结果表明提出的并联网络使得预测电场误差下降了18％以上，同时针对目标后向远场电场的预测，其相对均方根误差能够小于5%.

Automatic target recognition of SAR images based on global scattering center model

Based on the scattering center parameterized model and back propagation neural network, a far-field electric field prediction network for the target in full angles and wide frequency band is constructed in this paper. The fast prediction of the real and imaginary parts of the far-field electric field is realized by using the target position, amplitude, frequency and other data information through the network. Firstly, the position and intensity of the strong scattering points are extracted. Then, the two-dimensional angular domain and frequency domain are divided into regions. A parallel intelligent network architecture is constructed to establish the relationship between the scattering center parametric model and the high-precision far-field electric field. This method can reduce the frequency and angle dependence of the traditional scattering center model through the training of the novel parallel network. It realizes the rapid acquisition of the far-field electric field of the target. Because the influence of the scattering parameters in the existing model on the target electric field is fully used for reference in the process of network design, the neural network has clear physical significance and outstanding generalization ability. Compared with the traditional electric field reconstruction method based on geometrical theory of diffraction (GTD) model, this method has higher accuracy. The experimental results show that the proposed parallel network reduces the prediction error of electric field by more than 18%. At the same time, the error of the proposed parallel prediction network for the prediction of the target backward far-field electric field can be less than 5%.