DATA AUGMENTATION FOR HRRP BASED ON GENERATIVE ADVERSARIAL NETWORK

In radar automatic target recognition (RATR), the data-driven model has proven to be a powerful tool. Nevertheless, its high requirement on the training dataset becomes a major restriction in many critical applications, especially on high resolution range profile (HRRP) based RATR. Data augmentation methods can improve the capabilities of data-driven model by increasing data diversity. As a deep generative model, the generative adversarial network (GAN) has demonstrated the effectiveness in data augmentation. In this paper, the authors apply GAN to augment the training data in data-driven RATR based on HRRP. Specifically, a GAN with one-dimensional (1D) fully connected structure is trained on the acquired HRRP data to learn the distribution of the whole data set. Then, the resulting generator is used to generate artificial HRRPs as the supplement to the acquired HRRP data set. Finally, the RATR system is trained on the augmented data set. Experiment results show that the proposed GAN-based HRRP augmentation method outperforms traditional augmentation methods, such as time-shift and mirroring.