Self-supervised Speckle Reduction GAN for Synthetic Aperture Radar

In this work, we present a novel generative adversarial network (GAN) for speckle reduction in synthetic aperture radar (SAR) imagery that requires only knowledge of the noise statistics. Speckle is ubiquitous in SAR and can cause problems both for human interpretability and for automated processing such as automated target recognition. The speckle reduction GAN presented in this paper does not require image pairs to train on. Instead we take the output of a smoothing CNN, subtract it from the noisy input and compared the residual noise with simulated speckle drawn using well known SAR noise characteristics. This directly encourage the smoothing CNN to remove those image features that match our simulated speckle. For smoothing and discrimination network, we utilize a simple CNN with a small number of residual network blocks, and unchanging size from layer to layer. We train and show good performance on MSTAR radar data, as well as analyzing performance on a simulated test set. We will compare our results with other methods in the literature including the self-supervised deep image prior algorithm as applied SAR data.