Radar HRRP Data Augmentation Using CVAE with Extended Latent Space Distribution

Abstract

In this paper, we propose a variational autoencoder (VAE) based generative model with particular regard to the aspect angle sensitivity of the radar HRRP data of maritime vessels, and conduct high-resolution range profile (HRRP) data augmentation experiments to improve the recognition performance. Specifically, we train the extended conditional Variational auto-encoder (ECVAE) model to reconstruction data, and consider the latent space distribution of the sample as a more general multidimensional posterior Gaussian distribution. Discrete or continuous labels can be input to the model. Design a periodic latent distribution to deal with periodic labels. Use Kullback-Leibler (KL) divergence to evaluate the similarity of the distribution and reconstruct data with the latent space distribution which making the dimension as low as possible. Experiments based on MNIST data and measured vessels HRRP data show that the ECVAE model can augment the data of samples to improve recognition Performance, in especial in the case of a small number of data samples.