Despeckling is a longstanding topic in synthetic aperture radar (SAR) imaging. Many different schemes have been proposed for the restoration of SAR images. Among the different possible strategies, the methods based on convolutional neural networks (CNNs) have shown to produce state-of-the-art results on SAR image restoration. However, to learn an effective model it is necessary to collect a large number of speckle-free SAR images for training. To bypass this problem, we propose to directly use pre-trained CNN models on additive white Gaussian noise (AWGN) and transfer them to process SAR speckle. To include such CNNs Gaussian denoisers, we use the multi-channel logarithm approach with Gaussian denoising (MuLoG). Experimental results, both on synthetic and real SAR data, show the method achieves good performance.
Paper
Full text
SAR Image Despeckling Using Pre-trained Convolutional Neural Network Models
Semantic Scholar · Engineering · 2019
Abstract
Despeckling is a longstanding topic in synthetic aperture radar (SAR) imaging. Many different schemes have been proposed for the restoration of SAR images. Among the different possible strategies, the methods based on convolutional neural networks (CNNs) have shown to produce state-of-the-art results on SAR image restoration. However, to learn an effective model it is necessary to collect a large number of speckle-free SAR images for training. To bypass this problem, we propose to directly use pre-trained CNN models on additive white Gaussian noise (AWGN) and transfer them to process SAR speckle. To include such CNNs Gaussian denoisers, we use the multi-channel logarithm approach with Gaussian denoising (MuLoG). Experimental results, both on synthetic and real SAR data, show the method achieves good performance.