Despeckling Method for Single SAR Images Based on Zero-shot Learning
Speckle filtering is an important pre-processing step for synthetic aperture radar(SAR)image interpretation.In recent years,speckle-filtering methods based on convolutional neural networks(CNNs)have been rapidly devel-oped.However,supervised learning based methods lack speckle-free reference SAR images as ground truth,and self-supervised learning based methods rely on multi-temporal SAR images from the same scene for speckle filtering.How-ever,these additional auxiliary datasets are difficult to obtain in actual scenarios.In addition,self-supervised learning methods generally require large training datasets and deep networks for speckle filtering,resulting in high computational complexity.Therefore,a speckle filtering method for single SAR images based on zero-shot learning is proposed in this paper.The core idea of this method is to perform sublook decomposition on the test SAR image and select the paired sub-look images closest to the test SAR image.It is theoretically proven that using paired sublook images for self-supervised training of the network achieves the same filtering effect as using speckle-free reference SAR images for supervised training of the network.Therefore,the self-supervised loss function is designed to quickly train the lightweight speckle-filtering network,and the trained network can be used for filtering the test SAR images.Compared with the speckle-filtering methods based on supervised learning and self-supervised learning,the proposed method does not require speckle-free reference or multi-temporal SAR images for model training nor additional training data.Speckle filtering can be implemented by using any lightweight CNN.Experimental results on the Radarsat-2 and ALOS-2 datasets show that the proposed method reduces the parameters by 22 times compared to the reference method;thus,it can better sup-press speckles in homogeneous areas and preserve image details.
NETL
NSTL
万方数据
