The Synthetic Aperture Radar-based Image Augmentation Using Generative Adversarial Network

Synthetic aperture radars (SARs), characterized by their excellent performance, have found wide use in military applications. SAR-based image recognition is an important research area. Neural networks (NNs) have been widely used for image recognition and have demonstrated high recognition accuracy. However, NNs require a significant amount of data for training. In situations with limited training data, it is difficult to obtain satisfactory recognition accuracy using NNs. One approach toward alleviating the problems associated with small training sets is data augmentation. Here, we propose a novel SAR-based image augmentation method: using CycleGAN, aerial images are transformed into SARformat images, thereby allowing expansion of the SAR dataset. To improve the quality of these transformed images, we designed two filters: 1) a pre-filter for selecting images that are suitable for the transformation and 2) a post-filter for selecting suitable images after the transformation; this allows approximating the transformed images as real SAR images. We validated our method on the SSDD dataset. Our experimental results show that compared with the traditional data augmentation method, our method exhibits better recognition performance. With only 300 transformed images, the mean average precision improved from 66% to 76%, using only a small fraction of the SSDD dataset, thus overcoming the problem of small-sized training sets.