The Research Of SAR Image Generation And Data Expansion Based On Generative Adversarial Network

Synthetic aperture radar(SAR)plays an important role in the military and people's livelihood.It has been a research hotspot in the field of radar remote sensing for recent years.With the continuous development of deep learning technology,research methods for SAR image understanding and interpretation have also received new inspiration.However,research methods based on deep learning generally require sufficient labeled data.SAR images are expensive and difficult to obtain,which limits the application of deep learning techniques to SAR images.Therefore,research on the generation and expansion of SAR images has important research significance.With the research and analysis of the current state of the art,this paper attempts to use Generative Adversarial Networks to generate SAR images and expand data,and explores the significance of generated data for SAR image classification.This paper analyzes and introduces the characteristics of SAR images on the basis of Generating Adversarial Networks.The specific work includes the following aspects:First,SAR image generation method based on distribution and LBP-GAN.By analyzing the imaging mechanism and characteristics of SAR images that are different from optical images,SAR image information is introduced to the traditional Generative Adversarial Network.This paper proposes a distribution-based generation method,which improves the generation of noise potential space against the network,making it more consistent with the statistical distribution information of SAR images.In order to constrain the network's learning of SAR image texture features,an LBP loss function is proposed.The two methods respectively capture SAR image characteristics from low-level pixel features and high-level structural features,so that the generated image is closer to the real image,and has better generation effect and generation efficiency,which helps to improve the classification accuracy.Second,SAR image generation method based on structural similarity Generation Adversarial Network.Aiming at the problem that the generated result of the high-resolution SAR image is blur and contains noise,by improving the network structure,the Markov discriminator network is used to learn the regional details and texture features of the high-resolution SAR image,and a structural similarity loss function is proposed to ensure that the generated image and the real image structural similarity.By combining the hierarchical network structure of the pyramid,this method performs internal mixed mode learning of a single SAR image,and improves the overall effect of generating an image.This method further improves the similarity between the generated image and the real image,and at the same time improves the classification accuracy based on the generated expanded data.Third,SAR image generation method based on random mask and image completion.Aiming at the problem of insufficient SAR image generation diversity,the idea of image completion is proposed to generate images.Through the multi-component completion method and the use of mixed-mode learning between samples,the overall structural characteristics of the SAR image are captured.This method also improves the stability of the generation.A random mask method is designed to adapt to the structure of different features in the SAR image,which increases the randomness generated and diversifies the results.The expanded data has noise resistance,and can help to improve the classification accuracy.