Research On 3D SAR Data Enhancement Technique Based On Deep Learning

Synthetic Aperture Radar(SAR)plays a crucial role in military and civil applications with all-day,all-weather reconnaissance capability,and its ability to carry different platforms such as aircrafts and satellites.With the continuous expansion of SAR imaging applications,people have begun to explore in the field of 3D SAR imaging technology.In the imaging process,the mainlobe of the echo signal is weakened and sidelobe appears,while interference by noise,all of which affect the imaging quality and cause difficulties in subsequent target detection and other applications.With the development of deep learning technology,image processing networks have emerged and gradually expanded to the field of SAR images.However,the commonly used networks are mainly for 2D SAR images,and 3D networks are mainly applied to medical image segmentation and other fields.There are few deep network models for 3D SAR data processing,and it is difficult to perform well in data enhancement.In this thesis,we investigate the 3D SAR data enhancement techniques based on deep learning.The main research work is as follows:1.The basic principles of 3D SAR imaging and SAR data enhancement algorithms are described.Firstly,the imaging geometry model and echo signal model of 3D SAR are established.Secondly,the principle of 3D Back Projection(BP)algorithm is described,and the derivation of 3D imaging process of point target is made,and the simulation is carried out on MATLAB to verify the rationality of the algorithm.Finally,the current classical SAR data enhancement algorithms are introduced,and Finally,the current classical SAR data enhancement algorithms are introduced,and their performance and application scenarios are analyzed,and the advantages and shortcomings of each algorithm are evaluated.In addition,the evaluation indexes of data enhancement are introduced for subsequent objective analysis of the processed images.2.A three-dimensional SAR data enhancement algorithm based on CNN is proposed.Firstly,the principle and fundamental model of 3D CNN are introduced,the problems in the practical application of 3D network in image processing are described,and its feasibility in SAR data enhancement processing is analyzed.The process of constructing a training data set using MATLAB simulation is described,and a 3D CNN is constructed to enhance the 3D SAR data.The network is based on the 3D U-Net structure,using separable convolution to reduce the computational cost,adding skip connections to avoid the model degradation,and introducing residual learning to optimize the model performance in addition.3.A 3D SAR data enhancement algorithm based on GAN is proposed.Firstly,the basic principles of generative and adversarial networks and the problems such as instability during actual training are introduced.Based on the Wasserstein distance,an alternative to KL divergence and JS divergence.a generative and adversarial network model is built,while a gradient penalty is added to satisfy the constraints.To accelerate the training progress,a corrector is introduced and the loss function of the generator is corrected.The feasibility of the network model proposed in this paper for 3D SAR data enhancement is verified through the processing of real measurement data.