A Tailored cGAN SAR Synthetic Data Augmentation Method for ATR Application

This article proposes a method to simulate Synthetic Aperture Radar (SAR) targets for specific incidence and azimuth angles. Images synthesized by Electromagnetic Computing (EMC) are used to train a Conditional Generative Adversarial Network (cGAN). Two synthetic image chips of the same class and incidence angle, separated by two degrees in azimuth, are used as input to the cGAN. The cGAN predicts the image of the same class and incidence angle whose azimuth angle corresponds to the bisector of the two input chips. An evaluation using the SAMPLE dataset was performed to verify the quality of the image prediction. Running through a total of 100 training epochs, the cGAN converges, reaching the best Mean Squared Error (MSE) after 77 epochs. The results demonstrate that the proposed method is promising for Automatic Target Recognition (ATR) applications.