Remote Sensing Image Super-resolution Using Generative Adversarial Networks

In this paper,the Generative Adversarial Networks is proposed for the first time to be used in the Super-Resolution of Remote Sensing Images.The Generative Adversarial Networks is a very popular Generative Model in recent years.The basic principle is to estimate the parameters of probability distribution of the sample data under the premise of assuming that the sample data obey a certain distribution and use the idea of " zero-sum game" for reference,to establish two sub-models to improve their " ability" in mutual confrontation.In essence,it minimizes the distance between two sample probability distribution functions by choosing different measures,such as kullback-leibler divergence,Jensen-Shannon divergence and Wasserstein distance.Firstly,we review the development of Super-Resolution since it was put forward,especially for remote sensing images,which have been studied by those methods.Include four methods of obtain a high-resolution image through Image Interpolation,Statistical Analysis,Statistical Learning,and based on Reconstruction.The basic principle of Deep Learning is Local Perception and Weights Sharing,which is one of the main reasons for the great success of Deep Learning in many fields,mainly reflected in the image,voice,video,text and other tasks.This paper introduces the popular Deep Learning framework in recent years,which is the first toolkit for Deep Learning research.Then,this paper introduces in detail some kinds of classical models based on the idea of Generative Adversarial Networks and explains their advantages and disadvantages.finally,the experiments of generating confrontation network on face data sets are carried out,and good results are obtained.It also explains the specific idea of Generative Adversarial Networks in super-resolution for the Super-Resolution.When the High-Resolution images generated by the model are close to the HighResolution images in the training set,the probability distribution between them is close.Finally,we developed two experiments,one is based on Deep Convolution Neural Network in Remote Sensing images and non-remote sensing images Super-Resolution experiment,one is to Generative Adversarial Networks in Remote Sensing images and non-remote sensing images Super-Resolution experiment,in order to Generative Adversarial Networks in the Remote Sensing Image Super-Resolution effect of horizontal and vertical contrast.The development language used in this paper is python.based on the TensorFlow Deep Learning framework,the Super-Resolution with Convolution Neural Network model and the Super-Resolution with Generative Adverbial Networks model are developed.The metrics used include PSNR,SSIM,and RMSD to evaluate the effect of Super-Resolution.In addition,the effect can be visually observed from the generated image.