Semi-Supervised SAR ATR via Multi-Discriminator Generative Adversarial Network

As a supervised deep learning algorithm well-suited for image processing, convolutional neural network (CNN) has shown great potential on synthetic aperture radar (SAR) automatic target recognition (ATR) and achieved superior performance in recent years. However, the training of the deep convolution network depends heavily on sufficient labeled samples while the SAR images are scarce and difficult to obtain, and it is time-consuming to artificially annotate labels for raw images. In this paper, a semi-supervised recognition method combining generative adversarial network (GAN) with CNN is proposed. We generated unlabeled images with GAN and set them as the input of CNN together with original labeled images, so as to implement the effective training and recognition with limited training samples. In order to address the instability training issue caused by the adversarial principal of GAN, a dynamic adjustable multi-discriminator GAN (MGAN) architecture is introduced in the proposed framework. Meanwhile, the label smoothing regularization (LSR) is applied to regularize the semi-supervised recognition model of the CNN. Experiments carried out on the moving and stationary target acquisition and recognition (MSTAR) dataset have indicated that the proposed method possesses the ability to improves the accuracy and robustness of CNN system, especially when the training dataset is limited.