ISAR Image Generation of Ships by Inpainting Using SinGAN

Inverse Synthetic Aperture Radar (ISAR) images are important in civilian and military applications. This is particularly true in military applications, where they are used for surveillance. One of the applications that we focus on is the Automatic Target Recognition (ATR) system for military ships that processes ISAR data. Extensive ISAR datasets are required to develop robust target recognition systems for ships. In this paper, we propose the use of SinGAN, a deep learning model which is architecturally a Generative Adversarial Network (GAN), to generate ISAR images of ships using an inpainting technique. The methodology begins with a single ISAR image of a ship being trained with the SinGAN model. Then, clip-arts, outlines, or even photographs of ships are preprocessed to ensure optimal inpainting results. This paper also evaluates noise reproducibility and demonstrates dataset creation using the generated ISAR images. Finally, this dataset is used to train a custom Convolutional Neural Network (CNN) model.