Research on Few-sample Target Recognition Algorithm Based on GAN Network

With the development of high-resolution radar, modern radar echo signals contain target shapes, sizes, attitudes, movement patterns and so on. In recent years, deep learning has been widely used in target recognition and tracking. In this paper we identify four kinds of aircraft targets based on electromagnetic echoes of the targets by the LSTM with an attention mechanism. This model can capture key features, increase the weight of key features in the final classification, and improve the model recognition accuracy. The neural network training requires a large amount of data. However, the simulation of electromagnetic echoes of the targets is very time-consuming. To overcome this dilemma, we use Generative adversarial network (GAN) to generate training samples for data augmentation. It can significantly reduce the computation time for acquiring training samples and can improve the generalization ability of the model. The incident waves imping on targets are Gaussian pulses. Electromagnetic echoes of targets are calculated by the Finite Domain Time Domain (FDTD) algorithm in this paper. The training data comprises two parts. One is the simulated echoes of targets from different pitch angles and azimuth angles. The other is generated by the GAN network. It is demonstrated that the average recognition rate of the proposed method is better than the LSTM model without the augmented dataset.