CVGAN: A Cross-View SAR Image Generation Method for Enhancing the View Diversity

Abstract

ABSTRACT Although numerous classification algorithms have been successfully applied to synthetic aperture radar (SAR) automatic target recognition (ATR), the performance of most methods is limited by the lack of SAR images under different views. Considering traditional generative adversarial networks (GAN) generate fake SAR images same as the input, a novel cross-view GAN (CVGAN) is proposed to generate SAR images under disparate views for enhancing view diversity. In this network, a generator is first used to learn the mapping relations of image pairs to generate cross-view SAR images. Then, a pre-trained classifier is utilized to improve the view angle and category authenticity of the generated images. Finally, image pairs are fed into the discriminator to distinguish the generated SAR images from the real SAR images. To make the proposed network better learn the mapping relations of cross-view images, a modified loss function is extended to our network. The experimental results on the moving and stationary target acquisition and recognition (MSTAR) dataset demonstrate the effectiveness of the CVGAN on SAR image generation.