Research On Remote Sensing Image Deblurring Based On Wasserstein Generative Adversarial Networks

Remote sensing is an important means of Earth observation.Many key information extracted from remote sensing images has been widely used in reconnaissance,monitoring,prevention,early warning and other fields.In the process of remote sensing imaging,image blur due to factors such as long shooting distance,fast scanning speed,external light interference,atmospheric turbulence and large wide imaging greatly reduce the image quality.The main task of deblurring remote sensing image is to remove the blur in the remote sensing image while retaining its key features,so that remote sensing image can be applied to more actual fields.Most of the existing image deblurring algorithms are processed under the premise that the point spread function(blur kernel)and image noise of the blurred image are known.Even if the blur kernel and noise are unknown,they will be estimated and assumed by certain methods.However,in the actual imaging process,the reason for the blurring of remote sensing images is complex and versatile,and the blur kernel and noise distribution cannot be accurately known.For this reason,it is necessary to study the more accurate deblurring algorithm for remote sensing images when the image degradation factors are variable and unclear.This dissertation focuses on the deblurring problem of remote sensing image,and based on the deblurring model of Generative Adversarial Networks,the image deblurring is used to improve the quality of remote sensing image,including end-to-end remote sensing image deblurring model and lightweight remote sensing image deblurring model.The main researches and innovations of the thesis are as follows:1.A remote sensing image deblurring algorithm based on WGAN(Wasserstein Generative Adversarial Networks)is proposed.The goal of the algorithm is to train an end-to-end deblurred network model which is divided into two parts.One is the generator,designed as a residual block based convolutional neural network structure,used to perform deblurring processing to generate clear remote sensing images.The other part is the discriminator,which is used to discriminate between the deblurred image generated by the generative network and the original clear image.It is not only used to judge the true and false of the image,but also to assist the generative network to generate a deblurred image that is closer to the real image.To achieve nature awareness,weredesign the loss function by adding the content loss and the perceptual loss.Compared with the traditional method which estimating the blur kernel based on the image prior,our method has higher precision,wider range,more realistic effect and more adaptability.2.Based on depthwise separable convolution,a optimization algorithm of the remote sensing image deblurring network is proposed.The goal of the algorithm is to reduce the size of the model,improve the network efficiency,and make the network model more lightweight while keeping the previous model deblurring effect unchanged.Since there are many convolution operations in the previous model structure,replacing the general convolutions with the depthwise separable convolutions will greatly reduce the weight parameters of the network training and reduce the model training time.This method has some improvements on the overall visual effects and image quality evaluation indicators compared to other methods,and effectively reduces the size of the model,which makes the algorithm applicable to a wider range,or can be directly implemented in the embedded.