We fabricated an effective target recognition network-based on the convolutional neural network and the perturbation Shooting and Bouncing Rays algorithm for electrically large targets having varying geometrical shapes. The algorithm started by constructing a relationship between the local random variables and target's geometry by using the non-uniform rational B-spline surface modeling method. Inverse synthetic aperture radar (ISAR) imaging was performed quickly for perturbed geometrical shapes and the algorithm can achieve higher computational efficiency than the Monte Carlo method. The ISAR imaging database was generated, and the database can describe the local geometrical uncertainty of the target. All the ISAR imaging results were used to train the CNN-based training network. When an ISAR image of a target with an uncertainty shape was inputted, this target could be recognized via the proposed CNN-based target recognition network. Our numerical results proved that the recognition rate could go up to 90% even for complicated targets.

我们基于卷积神经网络和扰动射击和弹跳射线算法构建了一个有效的目标识别网络，用于具有不同几何形状的电大目标。该算法首先利用非均匀有理B样条曲面建模方法构建局部随机变量与目标几何形状之间的关系。逆合成孔径雷达（ISAR）成像对扰动的几何形状进行快速成像，该算法可以实现比蒙特卡罗方法更高的计算效率。生成ISAR成像数据库，该数据库可以描述目标的局部几何不确定性。所有 ISAR 成像结果都用于训练基于 CNN 的训练网络。当输入具有不确定形状的目标的 ISAR 图像时，可以通过所提出的基于 CNN 的目标识别网络识别该目标。我们的数值结果证明，即使是复杂的目标，识别率也可以达到 90%。