Multi-Discriminator Generative Adversarial Network for Semi-Supervised SAR Target Recognition

Convolutional neural network (CNN) has shown powerful potential on synthetic aperture radar (SAR) automatic target recognition (ATR). However, the training of deep structure of CNN has high requirement for sufficient labeled sample images while the existing SAR images are limited and difficult to obtain. In this paper, an improved recognition architecture which combines CNN with generative adversarial network (GAN) is proposed. We generate unlabeled images from the training dataset by GAN and use them together with the original images for SAR. Then the label prediction is implemented by the CNN based semi-supervised learning. In order to address the instability issue in training caused by the adversarial principal of GAN, the multi-discriminator GAN (MGAN) architecture is introduced in the proposed framework. Meanwhile, the label smoothing is utilized to regularize the semi-supervised model of CNN classification. Experiment results on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset indicate that the proposed method can effectively improve the recognition accuracy and robustness of CNN system.