Research On SAR Image Generation Models

Synthetic Aperture Radar(SAR)uses microwave imaging technology to obtain images,and has a wide range of applications in many fields,such as aerospace,astronautics and other fields.SAR image interpretation technology is a hot topic in the current SAR research field.However,in the research of SAR image interpretation technology,the problem of insufficient target samples in SAR images is often faced.Therefore,studying the SAR target images generation technology is an important direction to help complete the SAR image interpretation.The purpose is to increase the number of SAR target images,and help to better complete the task of SAR image interpretation.This thesis mainly focuses on SAR image generation models,including SAR image generation models based on Deep Convolutional Generative Adversarial Networks,the SAR image generation models based on Conditional Generative Adversarial Networks(CGAN),the angle information-guided SAR image generation method based on CGAN and Spectral Normalization for Generation Adversarial Networks(SNGAN),and the angle and category information-guided SAR image generation models under few sample conditions.The main work of this thesis is summarized as follows:1.The first part mainly studies the relevant theory of deep learning and generative adversarial networks.Firstly,the related theoretical knowledge of deep learning is introduced,including neurons and feedforward neural networks,convolutional neural networks,and back propagation algorithm.Then,the basic principles and training processes of Generative Adversarial Networks are studied.Finally,the basic principles and network models of the Deep Convolutional Generative Adversarial Networks are studied.2.Based on the above models,the second part mainly studies the angle information-guided SAR image generation models based on CGAN and SNGAN.Firstly,Conditional Generative Adversarial Networks is studied,including the basic model and application of CGAN,and the SAR image generation experiments are implemented.Then the basic model and principle of Wassertein Generation Adversarial Networks(WGAN)and SNGAN are studied in details.Finally,the angle information-guided SAR image generation method based on CGAN and SNGAN is studied,and the SAR image generation experiments are performed.The quality of the generated image is verified by classification experiment.3.The above models can obtain generated images with good quality when the training samples are sufficient,but when the number of training samples is limited,the generated images are not good enough.Therefore,this thesis further learns from the few-shot learning method and studies the SAR image generation models under few sample conditions.Firstly,the SAR target recognition method under few samples is studied,including the concept of meta-learning,experimental settings under few samples,and two SAR image target recognition methods based on meta-learning ideas,i.e.,the SAR image target recognition method based on the Siamese Network and the SAR image target recognition method based on the Relation Network.SAR image target recognition experiments are performed,and the results verify the effectiveness of the two SAR image target recognition methods under few samples.Then the meta-learning idea in the SAR target recognition method under few samples is used to improve the SAR image generation models.The network structure and principle of the angle and category information-guided SAR image generation models under few sample conditions are studied.Then two different network training methods are studied,including the traditional training method and the training method based on meta-learning.For two different training methods,SAR image generation experiments are performed,and the quality of the generated images is verified through classification experiments and similarity experiments.Finally,the target recognition experiment verifies that the generated images obtained by the studied model can help to improve the performance of SAR target recognition under limited samples.