Research Of SAR Image Data Diversity And Data Augmentation Method

In the research of SAR target recognition,the key consideration is the selection of data sets and how to obtain sufficient SAR image data.However,the existing data sets have low versatility.The right SAR image data is limited by the orbits and airspace when the SAR images are collected by domestic satellites.As a result,the images in many existing data sets are discrete in terms of observation angle and resolution.Many of the military targets that are missing,even in the category of confidentiality,are not available.These factors greatly limit the development of target recognition.This thesis focuses on the construction and expansion of SAR image datasets.The main contents include the research on the diversity and construction methods of SAR image data,the research on the generation of SAR image data and the method of directional expansion.The main work and contributions of this topic are:Firstly,by summarizing the application requirements of traditional SAR image datasets,analyzing the selection of target categories,determining the target selection scheme,a scientific and standardized data set diversity method is studied.Based on the proposed image data diversity method,the selection of image target categories,data preprocessing,classification,labeling,and multi-source auxiliary data are studied.Finally,a method for constructing a SAR image data set that can realize large-scale image data,is as close as possible to practical applications,has different coverage,is labeled,and is organized is established.Secondly,in order to make up for the defects of the existing data sets and enrich the SAR data,this thesis introduces and analyzes the current mainstream SAR image expansion methods.The SAR image generation method based on Deep Convolutional Generative Adversarial Networks(DCGAN)is studied.The quality of the generated image is initially evaluated by comparing a series of objective parameters with a linear composite image and a real image.The similarity between the real image and the gray histogram and the gradient histogram of the generated image is calculated,which further proves the potential for proves the potential of Generative Adversarial Networks in SAR image generation.Finally,in order to solve the problem that DCGAN is difficult to converge,the model is easy to collapse,and at the same time enrich and improve the angle and direction of the acquired SAR target image,Wasserstein GAN with gradient penalty(WGAN-GP)is introduced.An end-to-end model based on WGAN-GP for directional expansion of SAR images is proposed.The quality of the generated samples is evaluated by comparing a series of objective parameters with real samples.The validity of the expanded SAR image is verified by the recognition experiment(SOC condition and EOC condition),and it is confirmed that this method can enrich and improve the existing SAR data to a certain extent.