| Synthetic Aperture Radar(SAR)imaging has the advantages of all-weather and allday,so it is widely used in disaster monitoring,remote sensing mapping,and other fields.The SAR deep unrolled network imaging method unrolls traditional iterative processing methods into multi-layer neural networks with learnable parameters.It can simultaneously utilize massive existing SAR data and prior knowledge in traditional SAR imaging methods to fit complex SAR imaging mappings.It can reconstruct high-resolution SAR images with high precision from undersampled echo data,reducing the demand for data acquisition,storage,and transmission for high-resolution,wide-swath remote sensing mapping imaging.At the same time,it can also be applied to high-resolution imaging of small satellites,small drones,and other platforms,significantly expanding the range of SAR imaging,thus having broad application prospects in military and civilian fields.In this thesis,works including theoretical analysis,method research,and simulation experiments are carried out on the problems such as imaging problem solutions and network implementations in SAR deep unrolled network imaging method.The main contents of this thesis are as follows:(1)A non-sparse scene SAR imaging data set is constructed.SAR image samples of non-sparse imaging scenes are collected from the existing public remote sensing images.These images and the corresponding echo data generated constitute a SAR imaging data set for non-sparse scenes,which can meet the training and testing requirements of the SAR imaging networks.(2)A general iterative method for undersampled SAR image reconstruction is studied.According to the traditional model and typical solution of the undersampled SAR image reconstruction problem based on compressed sensing,an iterative reconstruction method using the generalized model of undersampled SAR image reconstruction is derived,which is the basis for the SAR deep unrolled network imaging methods.(3)A SAR unrolled network imaging method based on image features is proposed.The reconstruction model based on image feature constraints is designed to extract and utilize the prior information of SAR image features in non-sparse scenes.The iterative optimization process of SAR image reconstruction based on this model is then unrolled into a deep network.The image feature transformations and regularization functions conducive to SAR image reconstruction are learned from the training data to obtain better SAR image reconstruction performance in non-sparse scenes.(4)A SAR unrolled network imaging method based on prior distribution is proposed.The reconstruction model based on the prior distribution is designed for undersampled SAR imaging.Based on this,a deep unrolled network is constructed to utilize the prior distribution information of SAR images in the pre-trained generation network.The unrolled network can achieve high-precision SAR image reconstruction of non-sparse scenes and retain good reconstruction performance with limited samples of training echo data.In the above works,theoretical analysis and simulation experiments verify the proposed iterative solutions to the SAR imaging problem and the unrolled network imaging methods.The results show that the proposed method can effectively achieve highefficiency and high-precision imaging using undersampled SAR echoes. |
