A GAN-based Method for SAR Image Despeckling

摘要

Synthetic Aperture Radar (SAR) images are contaminated by multiplicative noise known as speckle. Most filtering methods are limited to noise statistics and requires complex parameter tuning to achieve the desired visual effects. To solve the above problem, a generative adversarial network (GAN) based method is proposed for SAR image despeckling. Firstly, homogeneous regions are selected manually and speckle samples are produced. Then, GAN is trained to learn the distribution of the speckle samples and generate the “realistic-looking” ones. Third, a convolutional neural network (CNN) is designed that specializes in removing the speckle. Experiments on simulated SAR images and real SAR images show the good performance of the proposed method with respect to both the visual effect and quantitative analysis.