Data Augmentation Method of SAR Image Dataset Based on Wasserstein Generative Adversarial Networks

The published Synthetic Aperture Radar (SAR) samples are not abundant enough, which is not conducive to the application of deep learning methods in the field of SAR automatic target recognition. Generative Adversarial Nets (GANs) is one of the most effective ways to generate image samples. In this manuscript, the gradient penalty method was used on the traditional DCGAN to solve the mode collapse problem. The improved model was used for data augmentation of the MSTAR dataset. All the experimental results show that the method was an effective way to generate SAR images with high quality.

I. Introduction

Synthetic Aperture Radar (SAR) is an effective way of ground observation. Automatic target recognition for SAR images is an important technology in the field of remote sensing image interpretation. At present, the SAR automatic target recognition technology based on deep learning has a large demand for SAR image data, but the limited number of public SAR image data disadvantaged the training of deep learning networks. Therefore, it is necessary to explore an effective SAR image data augmentation method to obtain enough SAR image data.