Despeckling of Synthetic Aperture Radar Image using Deep-Learning Model

This paper mainly studies despeckling method for Synthetic Aperture Radar (SAR) image. The denoising method based on deep learning usually needs a lot of samples and noiseless reference images to realize data regression. However, speckle noise is very different from random additive noise such as white Gaussian noise in that its distribution is affected by the background target and it is hard to obtain completely noiseless reference images. Therefore, we proposed a local learning method and constructed a deep learning model to overcome this challenge, which named Speckle-Denoise-Net (SD-Net). In proposed method, the speckle noise can be effectively suppressed without the need for noiseless reference images. It can not only solve the problem of lacking training samples and having excessive image size, but also take advantage of unsupervised learning for inherent noise to make a good balance between smoothing and structure-preserving compared with other methods. At present, the test results of this method (Including denoising effect, original information retention, and algorithm speed) have exceeded most of the existing traditional denoising algorithms and deep learning methods, which allow it to bring new inspirations to solve the problem of speckle noise.