ISAR Image Generation of Ships by Inpainting Using SinGAN

Inverse Synthetic Aperture Radar (ISAR) images is very important in civilian and military applications. Especially in the military where it is used for surveillance. One such application we focus on is the Automatic Target Recognition (ATR) system of military ships which gets 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 a Generative Adversarial Network (GAN) model to generate ISAR images of ships using inpainting technique. Firstly, a single ISAR image of ship is 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. Then, this dataset is used to train a custom Convolutional Neural Network (CNN) model.