基于零样本学习的单张SAR图像相干斑滤波方法

Despeckling Method for Single SAR Images Based on Zero-shot Learning

相干斑滤波是合成孔径雷达(Synthetic aperture radar,SAR)图像解译重要的预处理步骤。近年来,基于卷积神经网络(Convolutional neural network,CNN)的相干斑滤波方法得到了快速的发展。然而,基于监督学习的滤波方法缺乏无相干斑参考SAR图像作为真值,基于自监督学习的滤波方法大多需要同一场景的多时相SAR图像训练网络,但是这些额外的数据集在实际场景中较难获取。此外,自监督学习方法通常需要较大的训练数据集和较深的网络进行相干斑滤波,导致其计算复杂度较高。因此,本文提出了一种基于零样本学习的单张SAR图像相干斑滤波方法。该方法的核心思想是对待测试的单张SAR图像进行子视分解,选取与待测试SAR图像欧式距离最近的子视图像进行配对,理论上证明了使用配对的子视图像自监督训练网络能达到使用无相干斑参考SAR图像监督训练网络的滤波效果。因此,通过设计自监督损失函数快速训练轻量化相干斑滤波网络,将训练好的网络对待测试SAR图像进行滤波。相较于基于监督学习和自监督学习的相干斑滤波方法,本文所提方法不需要无相干斑参考或多时相SAR图像用于模型训练,也不需要额外训练数据,只需使用任意一个轻量化的CNN即可实现相干斑滤波。在Radarsat-2和ALOS-2实测数据上的实验结果表明,本文所提方法的参数量比对比方法低22倍,能更好的实现对匀质区域相干斑的抑制和图像细节的保护。

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 developed.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.However,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 sublook 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 suppress speckles in homogeneous areas and preserve image details.