A Method of SAR Target Data Augmentation Based on Component-Level Perturbation

Large-scale, diverse Synthetic Aperture Radar (SAR) image data plays a crucial role in Automatic Target Recognition (ATR). Directly acquiring SAR image data faces significant challenges in practical operation and acquisition costs, while methods such as electromagnetic simulation and image processing for obtaining image data also have insurmountable difficulties. Therefore, intelligent augmentation methods based on Variational Autoencoder (VAE) and Generative Adversarial Network (GAN) have been widely applied in the field of SAR image data augmentation. However, these methods are based on pixel-level augmentation, which not only deviates from the scattering mechanism of SAR targets but also tends to lose the structure and semantic information of the targets. To address this issue, we propose a component-based SAR image data augmentation method. We extract features from real SAR images, decouple the component information of the targets and perturb it, and finally obtain augmented data through attention mechanisms and reconstruction networks. This method not only reduces the adverse effects caused by background interference and occlusion on the targets but also helps to enhance the discriminability between different categories of targets, increasing the recognition rate of the targets on the dataset introduced in this paper by 2.56%.