A reconstruction method for obscured SAR images using VAE

In order to protect sensitive information, Synthetic Aperture Radar (SAR) images often need to be mosaicked or obscured, which brings difficulties to the further application of images. In this paper, a deep learning generation model Variational Autoencoder (VAE) is proposed to reconstruct damaged SAR images. First, the same type of normal SAR images are used for training, and then obscured SAR images are tested as input. The results show that this process effectively removes the mosaics or obscurations in SAR images and reconstruct the original visual effect of the images. It performs well in many evaluation indexes and realizes the reconstruction of obscured images.