Self-Supervised SAR Despeckling Using Deep Image Prior

Speckle noise produces a strong degradation in SAR images, characterized by a multiplicative model. Its removal is an important step of any processing chain exploiting such data. To perform this task, several model-based despeckling methods were proposed in the past years as well as, more recently, deep learning approaches. However, most of the latter ones need to be trained on a large number of pairs of noisy and clean images that, in the case of SAR images, can only be produced with the aid of synthetic noise. In this paper, we propose a self-supervised learning method based on the use of Deep Image Prior, which is extended to deal with speckle noise. The major advantage of the proposed approach lies in its ability to perform denoising without requiring any reference clean image during training. A new loss function is introduced in order to reproduce a multiplicative noise having statistics close to those of a typical speckle noise and composed also by a guidance term derived from model-based denoisers. Experimental results are presented to show the effectiveness of the proposed method and compare its performance with other reference despeckling algorithms. • A self-supervised learning method for SAR despeckling, denoted as S3DIP, is proposed. • The Deep Image Prior (DIP) approach is extended introducing a learnable noise matrix. • A histogram loss based on the knowledge of the speckle noise statistics is designed. • Existing model-based denoisers are exploited through the use of a guidance loss. • S3DIP outperforms both DIP and existing model-based denoisers on the intended task.