基于结构保持的星载SAR图像超分辨重构算法

Structure-Preserving Super Resolution Network for Spaceborne Synthetic Aperture Radar Images

星载合成孔径雷达(SAR)可以实现全天候、全天时、大幅宽的对地观测,但受天线尺寸和数据有损传输影响,高分辨率SAR图像的获取十分困难。针对该问题,提出一个基于结构保持的图像超分辨重构网络(SP-SRNet),以实现SAR图像从低分辨(LR)图像到高分辨(HR)图像的重构:使用一个轻量级深度卷积神经网络提取图像梯度图特征,为超分辨重构网络提供更多的结构信息;设计一个由像素损失和梯度损失组成的多目标函数优化SP-SRNet。利用ICEYE公司卫星高分辨SAR图像,采用双3次下采样算法构建SAR图像的LR-HR数据集,并在该数据集运用多种现有算法对比仿真验证。结果表明,提出的SP-SRNet在定量评估指标和主观视觉上均优于现有的超分辨重构算法。

Space-borne synthetic aperture radar(SAR)can realize all-weather,all-day,wide earth observation,but it is very difficult to obtain high-resolution SAR images due to the influence of antenna size and lossy data transmission.To solve this problem,an image super-resolution reconstruction network based on structure preservation(SP-SRNet)is proposed to reconstruct SAR images from low-resolution(LR)images to high-resolution(HR)images.A lightweight deep convolution neural network is used to extract image gradient features,which provides more structural information for the superresolution reconstruction network.A multi-objective function composed of pixel loss and gradient loss is designed to optimize SP-SRNet.The LR-HR data set of SAR image is constructed by using the satellite high-resolution SAR image of ICEYE company and bicubic downsampling algorithm.The results show that the proposed SP-SRNet is superior to the existing super-resolution reconstruction algorithm in quantitative evaluation index and subjective vision.