Statistical based CNN algorithm for SAR image despeckling

Despeckling is a key tool for the SAR image understanding and it is the first image pre-processing for other application such as classification, segmentation and detection. The importance of having a filter able to suppress noise without losing spatial details is fundamental. Among different approaches, nowadays several deep learning based algorithm for SAR despeckling are proposed. Most of them rely on simulated datasets based on certain hypothesis for the speckle that usually not fully exploit the characteristics of real SAR images, leading to methods that produce artefacts in areas with different characteristics from those present in the training. The aim of this work is to propose a multi-step despeckling process: in the first step a convolutional neural network trained under the fully developed speckle hypothesis with a statistical loss function is used for despeckling; later, by means of a statistical test and a ratio edge detector, the noise predicted by the network is used for detecting the not fully developed areas where the network will produce artefacts. Once this detection is done, an ad hoc filtering policy can be considered.