A Self-Supervised Iterative Convolutional Neural Network for Sar Image Despeckling

Le résumé fourni par la source

Speckle noise is an inherent noise in Synthetic Aperture Radar (SAR) images. This noise complicates the training of supervised methods, as clean SAR images are practically unattainable in real-world scenarios. Using simulated or multi-temporal images is a common solution, but this limits their efficiency in practical applications. To solve these problems, we propose a novel self-supervised learning method called SAR-IDR for SAR image despeckling. Firstly, we construct a mixed speckle noise model, and then add this noise to real SAR images to construct noiser-noisy image pairs. Secondly, noiser images are used as the input of the self-supervised neural network, and real SAR images are used as the first training target. Finally, the Iterative Data Refinement (IDR) is introduced to complete the iterative training of the network. The proposed method only needs to use real SAR images, making the image smoother while maintaining the point target intensity. We use the Radarsat-1 SAR images and the latest Haishao-1 SAR images as the experimental datasets. The experiment demonstrated the effectiveness of this method compared with existing methods.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.