Research On Multi-view SAR Image Target Data Augmentation Based On Generative Adversarial Networks

Synthetic Aperture Radar(SAR)has been widely used in military and civil fields.Among them,SAR image target recognition is an important part of many SAR applications.The main purpose of target recognition in SAR images is similar to that of optical images,which is to determine the target type accurately from unknown image data.However,due to the high cost of the equipments and the vehicles carrying those equipments for the acquisition of SAR images,public SAR image data is relatively less scarce than optical image data.Object recognition methods based on deep learning have significant effects in the field of image processing,but such methods require a large amount of data volume.In addition,due to the characteristics of SAR imaging technology,the target data of SAR images will show certain differences under the conditions of different azimuth angles and pitch angles.If the amount of SAR image data is insufficient,the recognition rate of target recognition model will be seriously affected when processing target data with different azimuth angles and pitch angles.Hence,if deep learning method is to be applied to SAR image target recognition,the problem of insufficient SAR image data must be solved.Aiming at solving the problem of how to effectively expand the target data of SAR image and to establish a complete data set of SAR image target,this thesis focuses on the research of multi-view target data augmentation methods of SAR images.The main work and contributions of this paper are listed as follows:1.On the basis of summarizing and comparing the existing generative models,the advantages of using GAN for data augmentation are analyzed,and determines two directions of data augmentation,namely data volume expansion and data enhancement.2.In order to solve the problems that low interpretability of generated images and insufficient precision of targets' azimuth angles when generating SAR image target data in the existing generative adversarial networks,information maximizing generative adversarial network is adopted to generate SAR image target data.By maximizing the mutual information between the latent code and the generated image,the target data of the SAR image with orderly change of azimuth angle is generated by changing the latent code value.At the same time,in order to improve the image quality of information maximizing generative adversarial network,we introduce the least square loss function to improve the loss function of information maximizing generative adversarial network.3.In order to solve the problems that the recognition rate is reduced due to the difference of distribution of the SAR image target data when one's type is the same but is under different pitch angle,the data augmentation method of the SAR image target data under different pitch angle is studied,and the AEGAN model based on the combination of the autoencoder and the auxiliary classifier generative adversarial network is proposed.The model can transform the distribution among different data domains,which brings the distribution among different data domains closer,and realize the distribution conversion of the same kind of SAR image target data under different pitch angles,so that data enhancement is achieved.In conclusion,a series of studies are carried out in this paper on data augmentation of multi-perspective SAR image target based on generative adversarial networks,which enriches the theory of target data augmentation of SAR image and provides an effective method for data augmentation of high-resolution SAR image target data under the conditions of different azimuth angles and pitch angles.