SAR Image Generation Using Structural Bayesian Deep Generative Adversarial Network

SAR image generation algorithms which can generate SAR images from different angles is of great significance in the field of radar automatic target recognition (ATR). Traditional SAR image ATR methods based on template matching need constructing a large number of targets' SAR image templates, and calculating the distance between the input SAR image and templates. However, the application of template matching methods is limited since no enough template images are generated resulting from the lack of the directly acquired SAR image data and the unguaranteed data quality. Although deep learning technology has made great progress in generating visible light images, it is also difficult to generate quantities of SAR images with high quality due to limited data acquisition means and complex imaging mechanism. More compact representation of the latent structure in the electromagnetic scattering data will be needed to fulfill the image quality and quantity requirements.In this paper, based on the deep learning technology, we propose a method to generate SAR image template of targets from multiple angles using the generative depth neural network.We use Bayesian network to construct a generative adversarial network (GAN), in which the parameters of the generative model and recognition model are pre-trained with general data sets to obtain a GAN model with the ability to generate SAR images from different azimuth angles. Then SAR images with different azimuth angles at the same elevation angle are used as samples. The model is re-trained to capture SAR images' features. Experiments show that this method can generate SAR images in new azimuth angles by using a small amount of data from other azimuth angles at the same elevation angle. This method can generate a larger dataset with a mount of high quality SAR image data for different targets, which can be used for template matching in multiple azimuth angles, as well as for other in-depth ATR study.