Self-supervised training strategies for SAR image despeckling with deep neural networks

Emanuele Dalsasso (1, 2, 3) , Loïc Denis (4) , Max Muzeau (5, 6) , Florence Tupin (1, 2, 3)

1 IP Paris - Institut Polytechnique de Paris
2 IDS - Département Images, Données, Signal
3 IMAGES - Image, Modélisation, Analyse, GEométrie, Synthèse
4 LHC - Laboratoire Hubert Curien
5 SONDRA - Sondra, CentraleSupélec, Université Paris-Saclay
6 DEMR, ONERA, Université Paris Saclay [Palaiseau]

Images acquired by Synthetic Aperture Radar (SAR) are affected by speckle, making their interpretation difficult. Most recently, the rise of deep learning algorithms has led to groundbreaking results. The training of a neural network typically requires matched pairs of speckled / speckle-free images. To account for the speckle present in actual images and simplify the generation of training sets, self-supervision approaches directly train the network on speckled SAR data. Self-supervision exploits a form of diversity, either temporal, spatial, or based on the real/imaginary parts. We compare the requirements in terms of data preprocessing and the performance of three self-supervised strategies.

Domains: Artificial Intelligence [cs.AI]; Neural and Evolutionary Computing [cs.NE]; Image Processing [eess.IV]

Emanuele Dalsasso, Loïc Denis, Max Muzeau, Florence Tupin. Self-supervised training strategies for SAR image despeckling with deep neural networks. 14th European Conference on Synthetic Aperture Radar (EUSAR), Jul 2022, Leipzig, Germany. ⟨hal-03589245v2⟩