Semi-Supervised Generative Adversarial Nets with Multiple Generators for SAR Image Recognition

Abstract

As an important model of deep learning, semi-supervised learning models are based on Generative Adversarial Nets (GANs) and have achieved a competitive performance on standard optical images. However, the training of GANs becomes unstable when they are applied to SAR images, which reduces the feature extraction capability of the discriminator in GANs. This paper presents a new semi-supervised GANs with Multiple generators and a classifier (MCGAN). This model improves the stability of training for SAR images by employing multiple generators. A multi-classifier is introduced to the new GANs to utilize the labeled images during the training of the GANs, which shares the low level layers with the discriminator. Then, the layers of the trained discriminator and the classifier construct the recognition network for SAR images after having been finely tuned using a small number of the labeled images. Experiments on the Moving and Stationary Target Acquisition and Recognition (MSTAR) databases show that the proposed recognition network achieves a better and more stable recognition performance than several traditional semi-supervised methods as well as other GANs-based semi-supervised methods.

Keywords

• Generative Adversarial Networks (GANs)
• semi-supervised recognition
• Synthetic Aperture Radar (SAR)
• deep learning